Add a standalone Qwen3 video-training example on WebVid-10M

[amazloumi]

Summary

• Reduces examples/qwen3_vlm/ to one working path: joint-decoder video captioning on WebVid-10M, started from pretrained components.
• convert_hf_backbone.py (new) — builds the VLM via build_vlm_wrapper (SigLIP2 from HF, fresh adapter), copies HF Qwen3 LLM weights into the backbone, and saves the whole model as one DCP [checkpoint].load_path reads. Honors tied embeddings; dense targets only.
• One config: vlm_qwen3_0.6b_joint_decoder_webvid.toml — avgpool (pool_window = 2) → 64 tokens/frame, so 18 frames × 64 = 1152 visual + 96 text = 1248 ≤ max_seq_len. LLM frozen; vision encoder + adapter train.
• Removes convert_multimodal_checkpoint.py (+ test) and the four arch configs — the new converter is dense/JD-only, so those paths weren't runnable here.
• README rewritten as a single walkthrough; configs use placeholder paths.
• No kempnerforge/ changes; main README / CHANGELOG / docs untouched.

Net: +494 / −1342. Validated end to end — a run from an HF-init checkpoint completed one WebVid epoch (loss 13.6 → ~2.7) and produced sensible captions on held-out validation clips.

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 (1676 passed)
• If distributed code changed — n/a
• If training loop / parallelism / optimizers changed — n/a
• uv run pytest examples/qwen3_vlm/tests/ passes (18 tests)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

🚀 New features to boost your workflow:

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