fix: Remove race condition during multi-GPU training

[uermel]

Fixes #27.

Summary

Under multi-GPU DDP, update_hparam performed an unsynchronized read-modify-write on the shared hparams.yaml, causing intermittent crashes and a corrupted hyperparameters file. This restricts the write to the global-zero rank.

Background

fit-model enables DDPStrategy when launched on >1 GPU (ddw/fit_model.py), running one process per rank. Each rank executes the full LightningModule lifecycle.

Call trace into the race

The hooks that reach update_hparam are not rank-gated, so every rank runs them:

on_train_start()        (epoch 0)         # all ranks
  └─ update_normalization()
on_train_epoch_end()    (every N epochs)  # all ranks
  └─ update_normalization()
       └─ update_hparam("unet_params", …)

All ranks resolve the same hparams.yaml (PL syncs the logger version across ranks → one shared version_N/ dir), then each performs:

hparams = yaml.safe_load(open(path, "r"))    # READ
hparams[key] = value                         # MODIFY
open(path, "w") ... yaml.dump(...)           # WRITE (truncates on open)

If one rank opens the file or is mid-write while another rank tries to read there is a race causing the error. Avoided using the global rank attribute of the module.