Config-driven Qwen3-0.6B VLM configs

[amazloumi]

Summary

• Make ModelConfig.head_dim an overridable field so the attention width can decouple from dim (Qwen3: dim=1024, head_dim=128); default dim // n_heads is unchanged
• Add qwen_patch_merger adapter (RMSNorm ln_q + 2-layer MLP)
• Thread head_dim through CrossAttentionBlock → CrossAttention so CA blocks share the backbone head_dim
• Add Qwen3-0.6B configs for all four arches: configs/train/vlm_qwen3_0.6b_{joint_decoder,cross_attn,mot,moma}.toml (SigLIP2 so400m-patch14-224 + qwen_patch_merger), identical backbone differing only in [vlm].arch
• Add tests/unit/test_qwen3_vlm_configs.py

Testing

• uv run ruff check kempnerforge/ tests/ passes
• uv run ruff format --check kempnerforge/ tests/ scripts/ passes
• uv run pyright kempnerforge/ passes (0 errors)
• uv run pytest tests/unit/ -v --timeout=60 passes (1560 passed, 12 skipped)
• If distributed code changed: uv run torchrun --nproc_per_node=4 -m pytest tests/distributed/ -v
• If training loop / parallelism / optimizers changed: uv run pytest tests/e2e/ --e2e -v

Closes #139
Refs: KEM-571

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
kempnerforge/config/model.py	100.00% <100.00%> (ø)
kempnerforge/model/adapter.py	97.63% <100.00%> (+0.35%)	⬆️
kempnerforge/model/cross_attention.py	100.00% <100.00%> (ø)
kempnerforge/model/transformer.py	94.62% <ø> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.