Spec1: dim-agnostic heads — fix MGFN hardcoded 2048 + text-align compatibility guard

[jinmang2]

Part of #16 (Spec 1 — Foundation).

Scope

• Fix MGFN hardcoded 2048 (src/models/mgfn/modeling_mgfn.py:83, x[:, :2048] / x[:, 2048:]) → derive magnitude split from config.feature_size. Other heads already read feature_size.
• requires_text_aligned: bool per head; pipeline factory rejects incompatible {backbone, head} (e.g. VadCLIP+VideoMAE) with a clear error. CLIP/InternVideo2 are text-aligned.

Done when

• MGFN runs on non-2048 features.
• Numerical-equivalence regression on the 2048 path passes (eager/fp32/TF32-off).
• Incompatible pair raises a clear error (unit test).

[jinmang2]

Consolidated into PR #22 (tracked in the PR checklist) to keep management to PR + Epic only. Code for done items is in PR #22; remaining items live in the PR's Spec-1 checklist.