Add format-aware FB conversion with GPU-safe parallelism

[weedmo]

What changed

• Route converter jobs by detected or requested input format (auto, mcap, or fb) and report skipped mixed-format recordings through job progress.
• Add a dedicated mcap_to_fb_convert job that rewrites legacy MCAP recordings into the Data Foundry FB layout using staging and atomic promotion.
• Run FB datasets through the data_foundry-conversion-worker GPU image via the Docker API while preserving cancellation and destination rollback behavior.
• Select FB converter concurrency from explicit GPU profiles:
  • RTX 5090: parallel=3, nvenc_parallel=3
  • RTX 4090: parallel=3, nvenc_parallel=2
  • Other GPUs, including RTX 5060-class cards: 1/1
• Keep FB_CONVERTER_PARALLEL, FB_CONVERTER_NVENC_PARALLEL, and FB_CONVERTER_GPU_MODEL as operational overrides.
• Extend the job enum, API types, Docker runtime dependencies, documentation, and regression coverage for the new paths.

Why

The previous FB wrapper always passed 2/2, so lower-memory GPUs inherited concurrency chosen for higher-end workers and could fail with GPU OOM. The converter also lacked a first-class distinction between legacy MCAP recordings and the Data Foundry FB input layout.

This change makes format selection explicit and gives lower-tier or unknown GPUs a conservative default while retaining tuned 4090/5090 profiles and manual overrides.

Impact

• Existing legacy MCAP conversion remains available through the same convert job.
• FB conversion requires the configured data_foundry-conversion-worker image and Docker socket access from the converter service.
• Deployments must rebuild/restart the converter image for the new runtime and profile selection to take effect.
• The database job enum gains mcap_to_fb_convert through idempotent schema compatibility statements.

Validation

• uv run pytest -q tests/test_converter_input_format.py tests/test_fb_converter.py tests/test_mcap_to_fb.py tests/test_mcap_to_fb_job.py tests/test_converter_queue_adapter.py tests/test_jobs_router.py tests/test_converter_dockerfile.py tests/test_ui_service_docker_ops.py tests/test_converter_mem_limit.py — 67 passed.
• npm run build — TypeScript and Vite production build passed.
• git diff --check — passed.
• RTX 5060 Ti profile smoke check selected parallel=1, nvenc_parallel=1.

Known validation gaps

• A real end-to-end FB conversion against the GPU worker image was not run as part of this publish step; the Docker lifecycle is covered with mocked regression tests.
• The full backend suite reaches an existing unrelated failure in tests/test_config.py: the test expects dataset_sources == ["lerobot", "lerobot_test"], while current settings also contain "raw". The first-failure run reported 37 passed and 19 skipped before that assertion.