Add proxy MSE training support

[XIAOEEN]

Summary

• add a ProxyModule that predicts white-box probability features from encoder hidden states
• add optional MSE supervision, proxy feature replacement during training, and proxy-only inference
• document proxy MSE training and evaluation flags in README

Checks

• PYTHONPYCACHEPREFIX=/Users/yangxiansen/Documents/PhantomHunter_FCS/.codex_work/pycache python3 -m py_compile main.py model.py dataloader.py utils.py scl_loss.py
• git diff --check

Note: a forward smoke test was not run because the local system Python environment does not have torch installed.