Unable to run GC6 coupled model at git_migration version of LFRic_Apps

### Version

main

### Are there any linked Issues or Pull Requests?

_No response_

### What happened?

[GC6](https://code.metoffice.gov.uk/trac/roses-u/browser/d/i/6/2/0/trunk) encounters sporadic failures with "rank 510 died from signal 6 and dumped core" which appear to come from NEMO when it is doing mpi_bcast after passing 1D variables into NEMO through OASIS. More information in [Dan's comment](https://github.com/MetOffice/lfric_apps/issues/196#issuecomment-3914061656).

### Relevant log output

```shell

```

### Original failure (now resolved)

Although the lfric_coupled rose stem test compiles and runs fine, we are currently unable to run any coupled GC6 workflows from the git_migration version of LFRic_Apps (we're still in the process of setting up building from github).

We appear to be seeing failures in the yaxt library in the job.err file. The last output in the PET files is:
`20260127100824.500+0000:P000:INFO : Physics: Initialised FD prognostic fields from UM2LFRic dump
20260127100824.503+0000:P000:DEBUG: Adding object on mesh [ 1] to inventory_by_mesh: nodal_rmultiplicity_w2_fv
20260127100824.504+0000:P000:DEBUG: Adding object on mesh [ 1] to inventory_by_mesh: u_lon_sample
20260127100824.504+0000:P000:DEBUG: Adding object on mesh [ 1] to inventory_by_mesh: u_lat_sample
20260127100824.505+0000:P000:DEBUG: Adding object on mesh [ 1] to inventory_by_mesh: u_up_sample`

Modifying the model run to be as close as possible to the rose-stem test indicates that the problem may be with having l_multigrid=true (the model got to timestepping when this was false but failed with a numerical error because other dynamics settings were incorrect).

Start with the GC6 coupled workflow linked in the issue and inspect job.err alongside the final PET output. Compare runs with l_multigrid=true and false, then trace the reported yaxt, mpi_bcast, OASIS, and NEMO failures. Done means the coupled GC6 workflow runs successfully from the git_migration version without the signal 6 failure.

Written by the indexing model from the issue text.