Validate job options before writing the job script

[Mohit-Ak]

run() currently does its option validation after it has already rendered and written the job script:

__generate_job_script(targets, case)
__validate_job_options()
__generate_input_files(targets, case)

__validate_job_options() raises for --no-mpi with more than one rank, nodes <= 0, tasks_per_node <= 0, or a malformed --email. Because it runs late, any of those aborts the run but leaves a freshly written <name>.sh in the case directory — and, since --clean and build() also run before it, a rejected invocation has by then already wiped the previous run's outputs and possibly triggered a build.

This moves __validate_job_options() to the top of run(), right after the targets are resolved and before the case file is loaded (loading executes case.py, so it's a side effect too). Validation only reads ARG(...), so the reorder has no other effect; valid invocations behave exactly as before.

Testing. Added toolchain/mfc/run/test_run.py, which drives run() under a sandboxed state.gARG with each rejected option combination and asserts that build, case.clean(), the job-script generator and the input-file generator were all not called, plus a control case that a valid invocation still reaches generation. On pristine master the four rejection tests fail (AssertionError: job script was written before option validation); with this change all five pass.

$ cd toolchain && python -m pytest mfc/run/test_run.py -q
5 passed
$ python -m pytest . -q          # the ./mfc.sh lint suite
375 passed, 8 warnings, 4 subtests passed in 8.68s
$ ruff check . && ruff format --check mfc/run/run.py mfc/run/test_run.py   # ruff 0.6.5
All checks passed! / 2 files already formatted

This is part (b) of #1511 only, per the issue's own suggestion to split it from the (a)/(c)/(d) tidy-ups.

Fixes #1511