Restore KVzip forward after context errors 🤖🤖🤖

[fallintoplace]

Summary

• Restore model.model.forward from an inner finally around the KVzip context yield.
• Preserve the normal path where the original forward is restored before KVzip's post-yield compression/scoring phase.
• Merge the branch with the latest NVIDIA:main and drop the extra test file per maintainer feedback.

Root cause

KVzipPress.__call__ monkey-patches model.model.forward before yielding to the caller, but the original method was restored only after a normal return from the with block. If the caller raised inside the context, the outer cleanup removed hooks and reset internal state without restoring forward, leaving the model patched after exit.

Validation

• make style
• UV_PYTHON=3.12 uv run pytest tests/presses/test_head_compression.py -k "KVzipPress"

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[SimJeg]

/ok to test 6f2012f

[SimJeg]

Hi @fallintoplace, thanks for your contribution. Could you merge your branch with main so that we can run tests ? Also for this PR, the additional test is not needed