fix(mistral): freeze text-only visual branches and enable resume

[yuhezhang-ai]

What does this PR do?

The Ministral3 SQuAD recipes adapt only the language model, but full fine-tuning still includes the unused vision tower and its sibling multimodal projector in Adam. Freeze both explicitly with the existing module-selector config and re-enable checkpoint-resume validation for both full and PEFT training.

PR #3635 already materializes missing Adam state before saving new checkpoints. Explicit freezing additionally keeps permanently inactive visual parameters out of the optimizer. PEFT already freezes these branches through its base-model policy and LoRA exclusions; the explicit selectors document and enforce the text-only recipe policy in both examples.

Changelog

• Set freeze_config.freeze_modules to model.vision_tower and model.multi_modal_projector in both Ministral3 SQuAD recipes.
• Remove both ci.checkpoint_robustness.skip_resume exemptions and their stale workaround comments.
• Add a tiny real-VLM regression test that reads each recipe and verifies frozen visual parameters, exclusion from optimizer groups, and retained trainable language/LoRA parameters.

Validation

• 48 focused CPU freeze/selector/optimizer-state tests passed.
• Four tiny-VLM CPU strict DCP optimizer round trips passed: full and PEFT, each with and without explicit freezing. Restored optimizer state and the next update matched exactly. This does not substitute for distributed 3B recipe coverage.
• Repository Ruff formatting/lint and git diff --check passed.
• Generated CI definitions include resume for both recipes and retain cross_tp_reload for full fine-tuning.
• Scoped EOS CI passed for commit df3476d0f5c05a1fbe407b5371b14790aef71e24: pipeline 69061370. Both recipes ran on 8 H100 GPUs with the shipped CI overrides and resume enabled.

Recipe	Source reference/parity	Train/save	AutoModel/HF reload	Resume	Cross-TP reload
Full	PASS	PASS	PASS	PASS (67 s)	PASS
PEFT	PASS	PASS	PASS	PASS (76 s)	Not configured

Before this PR is ready for review

• Read and followed the contributor guidelines; commit includes DCO signoff.
• Added the necessary focused tests.
• Documented language-only adaptation in the recipe comments.
• Full and PEFT scoped checkpoint-robustness CI passes, including resume.

Related to AMINT-9.

[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.

[yuhezhang-ai]

/ok to test df3476d

[athitten]

Thank you @yuhezhang-ai !