Separate out MPI and multi-GPU examples

[jwallwork23]

Closes #257.
Closes #253.
Note merges into the branch for PR #222.

Checklist

• Remove MPI from example 3.
• Test revised example 3 on two GPU devices.
• Write CPU-only MPI example.
• Add MPI example to test suite.

[jwallwork23]

The new MPI example is working fine locally for me, but crashes on the CI.

See

• https://github.com/Cambridge-ICCS/FTorch/actions/runs/12995413832/job/36242065764?pr=258
• https://github.com/Cambridge-ICCS/FTorch/actions/runs/12995136703/job/36241158041?pr=258 (previous push that didn't have comm size > 1 check)

The Windows workflow will need MPI installing, too:
https://github.com/Cambridge-ICCS/FTorch/actions/runs/12995413835/job/36242065775?pr=258

[jwallwork23]

The order of the examples is no longer particularly aligned in terms of increasing complexity. We should follow this PR by reordering them to give a better flow when read in series. Perhaps
simplenet -> looping -> resnet -> multi I/O -> MPI -> multi-GPU -> autograd.