feat: fix transformers 5.x compat, dtype mismatch, add wandb logging

- Fix post_init() always called (required by transformers >=5.x for
  all_tied_weights_keys)
- Fix dtype mismatch: cast fusion layers and heads to float32 after
  from_pretrained loads everything as bfloat16
- Add wandb logging (project/entity/run-name) to training loop
- Add wandb config fields and CLI args