fix(harness): install declared submission packages

[shravandoda]

Summary

• Install a submission's pinned extra_pip_packages in the Modal image.
• Keep prefetch on the frozen base environment.
• Install submission dependencies while building the image, before timed training begins.

Why

TASK.md permits pinned open-source packages and submissions/TEMPLATE/config.yaml
documents extra_pip_packages, but the Modal verifier currently ignores that field.
Submissions using an allowed package therefore fail before training.

Validation

• python -m py_compile harness/modal_verify.py
• Existing prefetch behavior remains unchanged.

[Saivineeth147]

Good catch, this field was documented but never wired up. Reviewed the diff and the approach is right, but one change needed before merge: build two images, not one. The train sandbox can use the augmented image, but eval_run has to stay on the frozen env.lock image. Otherwise a submission's pip package can patch the evaluator at import time (sitecustomize, a .pth shim, or shadowing transformers/datasets) and rig its own score. The fresh eval sandbox is our anti-tampering boundary, so nothing submission-controlled should be in it.

Concretely: keep image = build_image(modal, sub_cfg...) for train_run, and pass a separate build_image(modal) (no extras) into eval_run.

Two smaller things while you're in there: reject anything in extra_pip_packages that isn't pinned with ==, and print the installed list into the run log so it lands in the verification report.

Update this and I'll merge same day, it unblocks your #9.