[4] Clean train config by using hydra config overriding

[vinamarora8]

Hydra allows us to have a base config and override it with another config on top. We can use this for POYO where we have a mostly stable base config, and each different training target overrides/adds only some fields. Having a base config means we can add default config fields later without breaking any dataset-specific configs.

Also, use of proper hierarchies in base config cleans it up. Example:

optim:
  base_lr: 3.125e-5  # scaled linearly by batch size
  weight_decay: 1e-4
  lr_decay_start: 0.5  # fraction of epochs before starting LR decay

wandb:
  enable: true
  entity: null
  project: poyo
  run_name: null
  log_model: false

This PR adds base.yaml which is inherited by train_poyo_mp.yaml, train_allen_neuropixels.yaml, train_mc_maze_small.yaml. Does not update other configs as they don't work as it is with current code.