[Code scan] Forward MolTrainHF extra training options into config

This issue is a result of a Codex global repository scan.

## Summary
MolTrainHF accepts **params but never applies them to the saved config. Options consumed by HFNNModel and the UniMol data/training stack, such as freeze_layers, freeze_layers_reversed, loss_key, data cleaning options, transfer-learning paths, and conformer cache controls, are silently dropped.

## Code references
https://github.com/deepmodeling/unimol_tools/blob/4596596aa8f73eb462d5cc5a921d79966d0465da/unimol_hf/trainer.py#L31-L41
https://github.com/deepmodeling/unimol_tools/blob/4596596aa8f73eb462d5cc5a921d79966d0465da/unimol_hf/trainer.py#L48-L106

## Impact
Users can pass supported training options to MolTrainHF and get no error, but the options have no effect. The saved config also omits them, so prediction or resumed workflows cannot reproduce the intended setup.

## Suggested fix
Add explicit parity arguments for supported MolTrain options or merge validated params into config after known fields are set, then persist them in config.yaml.