lfric_atm_nwp_gal9 loses bit comparison across NRUN-CRUN boundary

### Version

main

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

#193 #202

### What happened?

There is an NRUN-CRUN test for the lfric_atm_nwp_gal9 C12 configuration, but the CRUN runs for only one time-step. This is insufficient to show up two causes of divergence. After dealing with these two issues, bit comparison across an NRUN-CRUN boundary was demonstrated for a longer run.

1. rh_crit is reinitialised to the configuration value at the start of the CRUN. But this influences only radaer which runs at the end of the time-step, meaning that the divergences do not affect the fields in the checksum file.
2. dust_mrel and soil_sand are fields initialised from ancillary files. But the ancillary files are not read in by CRUNs. Note, that in the C12 test the initial problem affected only a single point of the mesh.

@iboutle has proposed and tested a fix for item 1 above, but it is not yet on trunk (#193). Essentially, radaer will always use the value of rhcrit from the configuration namelist as is done in the UM:

https://code.metoffice.gov.uk/trac/lfric_apps/log/main/branches/dev/ianboutle/vn2.2_radaer_bitcomp

For item 2, some further analysis is required by relevant science section owners/experts to check on which fields should be checkpointed for which glomap configuration options. For example, the dust_srce routine also takes soil_clay as an input but the longer NRUN-CRUN test passed even when soil_clay was not checkpointed. The fact that no issue arose may simply be due to chance (noting that the soil_sand issue only affected a single point). (#202)

Once these are on trunk, tests should be expanded.

### Relevant log output

```shell

```