Basic fixes for the trends diagnostics

Context

Branches impacted: main
Dependencies: TRD

Analysis

The trends diagnostics seem to have been broken since at least the summer body development (16443a0b). For the NEMO 5 release, they should at least be able to run in an ocean-only configuration.

The code is broken due to the following issues:

Some of the trends diagnostics subroutines have incorrect array argument declarations

Most of the trends diagnostics subroutines use assumed shape syntax for the input array declarations, which will lead to incorrect results for arrays with less than 2 halo points.

For example, trd_pen is called via trd_tra_mng. These subroutines both use assumed shape syntax for their input array declarations:
```
REAL(wp), DIMENSION(:,:,:), INTENT(in) ::   ptrdx, ptrdy   ! Temperature & Salinity trends
```
```
REAL(wp), DIMENSION(:,:,:), INTENT(inout) ::   ptrdx   ! Temperature or U trend 
REAL(wp), DIMENSION(:,:,:), INTENT(inout) ::   ptrdy   ! Salinity    or V trend
```
However, trd_tra_mng is called using input arrays with no halo points:
```
ALLOCATE( trdtx(T2D(0),jpk), trdty(T2D(0),jpk), trdt(T2D(0),jpk), avt_evd(T2D(0),jpk), STAT= trd_tra_alloc )
!
REAL(wp), DIMENSION(T2D(0),jpk) ::   ztrds
!
CALL trd_tra_mng( trdt, ztrds, ktrd, kt, Kmm )
```
The indices of trdt/ztrds will start from 3, but due to the assumed shape declarations in trd_tra_mng and trd_pen the indices for ptrdx/ptrdy will start from 1. This will result in the wrong part of the arrays being accessed by trd_pen, leading to incorrect results.
XIOS grid definitions are incorrect for most tracer trends diagnostics

At present, most of the trends diagnostics in the tracer section of field_def_nemo-oce.xml are defined on the grid_T_3D grid. However, the arrays passed to trd_tra_iom (via trd_tra_mng) have no halo points, which results in XIOS crashing.

bu and bv are accessed outside of array bounds in trd_ken

bu and bv are declared without halo points:

ALLOCATE( bu(T2D(0),jpk) , bv(T2D(0),jpk) , r1_bt(T2D(0),jpk) , STAT= trd_ken_alloc )

but are accessed on halo points:

DO_3D( 0, 0, 0, 0, 1, jpkm1 )
   zke(ji,jj,jk) = 0.5_wp * rho0 *( uu(ji  ,jj,jk,Kmm) * putrd(ji  ,jj,jk) * bu(ji  ,jj,jk)  &
      &                           + uu(ji-1,jj,jk,Kmm) * putrd(ji-1,jj,jk) * bu(ji-1,jj,jk)  &
      &                           + vv(ji,jj  ,jk,Kmm) * pvtrd(ji,jj  ,jk) * bv(ji,jj  ,jk)  &
      &                           + vv(ji,jj-1,jk,Kmm) * pvtrd(ji,jj-1,jk) * bv(ji,jj-1,jk)  ) * r1_bt(ji,jj,jk)
END_3D

Fix

In principle, the arrays input to trd_tra and trd_dyn do not need to be defined on halo points for all diagnostics. They can generally be declared with shape DIMENSION(T2D(0),jpk), except:

In trd_tra, where trd_tra_adv (tracer advection trend components) requires the input arrays to be defined over loop bounds (1, 0, 1, 0) for the horizontal components
In trd_dyn, where trd_vor_zint requires the input arrays to be defined over loop bounds (0, 1, 0, 1) and trd_ken requires them to be defined over loop bounds (1, 0, 1, 0)

The issues described above can mostly be addressed by reducing the shape of these arrays and ensuring they are defined on the required halo points:

All arrays passed to trd_tra have been declared with shape DIMENSION(T2D(0),jpk), except for those used for the horizontal tracer advection trend components (jptra_xad/jptra_yad) which have been declared with shape DIMENSION(T2D(1),jpk)
All arrays passed to trd_dyn have been declared with shape DIMENSION(T2D(0),jpk). In trd_dyn, input arrays for the trd_ken and trd_vor diagnostics are copied to working arrays with shape DIMENSION(T2D(1),jpk) which are then passed to lbc_lnk. This lbc_lnk replaces the calls in trd_ken and trd_vor

and adjusting the TRD subroutines accordingly:

A wrapper function has been added to trd_tra/trd_tra_adv so that they correctly handle varying input array shapes

This approach is used elsewhere in the code, e.g. in eosbn2.F90
In other subroutines called from trd_tra, the input array declarations have been changed to DIMENSION(T2D(0),jpk)
Input array declarations in trd_dyn and its called subroutines have been changed to DIMENSION(T2D(0),jpk), except in trd_ken/trd_vor/trd_vor_zint* where they have been changed to DIMENSION(T2D(1),jpk).

Additionally:

The XIOS grid definitions for the tracer trends diagnostics have been corrected:

- <field_group id="trendT_odd"  grid_ref="grid_T_3D">
- <field_group id="trendT_even" grid_ref="grid_T_3D">
+ <field_group id="trendT_odd"  grid_ref="grid_T_3D_inner">
+ <field_group id="trendT_even" grid_ref="grid_T_3D_inner">

The out-of-bounds array accesses in trd_ken have been resolved by declaring bu and bv with DIMENSION(A2D(1),jpkm1) and defining them on the required halo points

Other changes

Fixed a divide by zero for the jptra_zdfp case in trd_tra:

- z1e3w = 1._wp / ( e3w(ji,jj,jk,Kmm) * tmask(ji,jj,jk) )
+ z1e3w = ( 1._wp / e3w(ji,jj,jk,Kmm) ) * tmask(ji,jj,jk)

Wrapper functions were also added to trd_trc/trd_mxl_trc_zint, but these changes were made mainly because trd_trc is called from trd_tra and they have not been tested.
Removed some unnecessary uses of T2D(0), e.g. where z2d(:,:) would be equivalent to z2d(T2D(0))
Replaced uses of T2D in module-level array declarations with A2D

This has no impact since the tiling is disabled for trends diagnostics, but it is more consistent with the planned coding conventions for NEMO 5 (T2D for local working arrays, A2D for global and module-level arrays)
Some other minor optimisations

Other issues that aren't fixed here

Several namtrd options are disabled by ctl_stop calls
- ln_tra_mxl/ln_vor_trd- 'ML tracer and Barotropic vorticity diags are still using old IOIPSL'
- ln_PE_trd- 'trd_pen_init : PE trends not coded for variable volume'
- ln_dyn_mxl- 'ML diag on momentum are not yet coded we stop'
The wind stress trends are not restartable when using RK3

With RK3, ketrd_tau (trd_ken) and utrd_tau/vtrd_tau (trd_dyn_iom) are not restartable because they depend on utau_b/vtau_b, which are only updated at the first timestep.

@techene - I haven't fixed this because RK3 doesn't seem to use utau_b/vtau_b and I'm not sure what to replace it with. Perhaps something to add to the list for when TRD is rewritten?
Calls to trd_tra are commented out in tra_adv_cen and tra_adv_mus

These subroutines were rewritten in 16443a0b so that the trends arrays are now 2D instead of 3D. At present, trd_tra can only handle 3D input arrays.
The passive tracer trends diagnostics may still be broken

I haven't done any testing with passive tracers, so I don't know if these trends diagnostics are functional.

To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information