Multimodal: Add Cross-Attention Arch

[amazloumi]

Summary

Adds the Cross-Attention VLM arch (arch = "cross_attention", Llama-3-V style) on top of the VLM foundation that landed in PR #76 . The residual stream stays text-only; image features flow as K/V into separate CrossAttentionBlocks inserted at a configurable cadence. CA blocks are zero-initialized so adding the arch on top of a text-only checkpoint is identity at step 0 and learns from there.

This PR stacks on PR #76, so the diff shows only the CA additions.
_RESERVED_ARCHS shrinks to ("mot",) — Mixture-of-Transformers will land in the next PR.

Testing

• ruff check kempnerforge/ tests/ scripts/ — clean
• ruff format --check kempnerforge/ tests/ scripts/ — clean
• pyright kempnerforge/ — 0 errors / 0 warnings
• sphinx-build -W --keep-going -b html docs docs/_build/html — clean (no Sphinx warnings)
• pytest tests/unit/ tests/integration/ --cov --cov-branch — 1130 passed, 1 skipped, 81% coverage
• torchrun --nproc_per_node=4 -m pytest tests/distributed/ --slow — 81 passed on 4× H200 (full distributed suite incl. 6 JD + 8 CA VLM tests)
• CA + MoE smoke (tests/distributed/test_vlm_cross_attn_fsdp.py::TestMoEWithVLM) — passes on 2-GPU FSDP2

Follow-ups

• PR 3 — Mixture-of-Transformers arch. MoTConfig + MoTBlock + MoTStrategy + JD→MoT warm-start helper. Per-modality Q/K/V/O + per-modality FFN at every layer, single global SDPA. Reserved in _RESERVED_ARCHS here.