Add web app layer: FastAPI backend, job queue and per-frame coordinate CSV export

[dylangriffinshub]

Adds a deployable web app on top of the existing CLI, plus the per-frame coordinate export the pipeline did not have.

The CLI (examples/soccer/main.py) is untouched. Everything new lives in a separate pitchvision/ package.

Pipeline (pitchvision/pipeline.py)

• Extracts the tracking flow into a reusable run_pipeline() that a server can call, rather than a generator tied to a CLI and local file paths.
• Adds a per-frame CSV export: one row per tracked object with frame, timestamp_s, tracker_id, class, team, centre point, bbox and confidence.
• Caches team assignment per tracker id. A player's shirt does not change, so each track is classified once instead of on every frame. This is the main speedup: SigLIP embedding was previously repeated for every detection on every frame.
• PipelineConfig exposes max_seconds, frame_stride, imgsz, device and torch_threads so the same code runs on a small CPU box or a GPU.

Web app (pitchvision/api.py, pitchvision/static/)

• FastAPI service with a SQLite-backed job queue and a single background worker, so concurrent uploads serialise instead of thrashing the machine.
• Upload page with progress polling; OpenAPI docs at /api/docs.
• Jobs interrupted by a restart are marked failed on startup rather than left hanging in processing.

Bug worth fixing regardless of this PR

TeamClassifier requires sentencepiece and protobuf (SigLIP's processor pulls them in). Neither is in examples/soccer/requirements.txt, so on a clean install constructing the classifier raises ImportError.

That failure mode is quiet and damaging: if the exception is caught broadly, clustering is skipped and every player is labelled team 0, so the run looks successful while producing wrong data. This PR surfaces the reason via PipelineResult.team_error instead of failing silently. You may want the requirements.txt fix upstream on its own.

Verified

On a 6s broadcast clip: 432 / 426 team split, referees detected separately, ball tracked, 956 CSV rows across 50 processed frames. Sample input, annotated output, stills and CSV are committed in docs/samples/.

Performance

Measured on an Intel i3-8100B (4 cores, no GPU): 1.20 s/frame at imgsz=640, 2.67 s/frame at 960. CPU inference is the bottleneck, which is why the defaults cap clip length and stride. On CUDA, PipelineConfig(device='cuda') with frame_stride=1 is the only change needed.

Known issue, pre-existing

Goalkeepers are assigned by proximity to team centroids. Goalkeeper kit matches neither outfield team, so they can be miscoloured. Behaviour inherited from resolve_goalkeepers_team_id, not introduced here.

Notes

• .gitignore now excludes var/ (runtime state) and the downloaded weights.
• Four <video> tags in the README were replaced: they point at Git LFS files not present in the repo, and GitHub does not render relative <video src> anyway, so they showed as broken. They now reference the committed samples.
• MIT licence and upstream attribution to roboflow/sports retained.

Happy to split this up or rework any part of it.