Aims to resolve #38

• Adds --fine-tune option to janus train

  • Given the current differences, this feels more natural than an entirely different command
  • Currently only checks that foundation_model is in the config for fine-tuning, and not present otherwise, and that the value of foundation_model is either a valid MACE-MP/MACE-OFF string or a path that exists
  • We could load the model (if given a path, or even if given a valid string), but we would be unable to pass this directly to MACE

• Adds tests based on data and configs from @alinelena in add fine tuning #38

  • Modifications include:
    • File names and locations for train/test/validation data
    • max_num_epochs reduced from 10 to 1 for both configs
    • Uses mace_mp_small.model model for fine-tuning, as we use it elsewhere
    • Changes default_dtype to float64 to match model for fine-tuning
  • Skipped if unable to load run_train from MACE, so not currently tested in CI, but can be tested localled
  • The fine-tuning test is quite slow (~1 minute) - are there ways to reduce this without inhibiting the function of the test? It may be inevitably given the task