feat(annotations): export COCO/YOLO 2D annotations from a scene manifest

[dchaudhari7177]

What & why

The manifest already records, for every named point of every entity at every frame, each camera's pixel projection (uv) plus the in_view / visible flags — exactly the ground truth a 2D detection + keypoint training set needs. This adds a typed exporter so a simulated scene doubles as a synthetic training-data generator, not just a triangulation benchmark.

Closes #40.

The exporter (multicam_sim/annotations.py)

• export_coco(manifest) -> CocoDataset / write_coco(manifest, path) — COCO keypoint-detection JSON.
• export_yolo(manifest) -> YoloDataset / write_yolo(manifest, out_dir) — one YOLO(-pose) .txt per image + a classes.txt.

Conventions:

• One image per (camera, frame) pair (file_name = cam{id}/frame_{frame:06d}.jpg, a logical reference since the sim doesn't render pixels).
• One annotation per entity that has at least one in_view point on that image.
• Bounding box = the axis-aligned pixel bounds of that entity's in_view points, so the box always lies within the camera frame (off-image points are excluded).
• Keypoints follow the entity's point set with the COCO visibility flag derived from the manifest: 2 = in_view and visible, 1 = in_view but occluded, 0 = not in_view (absent, x = y = 0).
• Categories are entity ids; the per-entity skeleton (edges) becomes the COCO skeleton (1-indexed pairs into that category's keypoints).
• YOLO coordinates are normalised to [0, 1] by each image's width/height.

Everything is typed all the way through with pydantic models, mirroring manifest.py. Nothing here depends on an optional extra.

Tests (tests/test_annotations.py)

• Round-trip (the issue's ask): export → parse the YOLO line back → denormalise → boxes match the manifest uv within tolerance; COCO bboxes are checked against the manifest uv directly.
• Visibility-flag mapping tracks in_view / visible (the pose scene's occluded left_wrist on camera 1 exercises the 1 flag; a synthetic manifest exercises the 0/absent path and that it's excluded from the bbox).
• Image coverage (every camera×frame pair, unique ids), category/skeleton structure, multi-entity categories, and that write_coco / write_yolo emit parseable files.

Verification

• New tests: 9 passing.
• Full collectable suite green; ruff check, ruff format --check, and mypy --strict (28 src files) all clean.
• The 5 pre-existing collection errors in the suite are unrelated (tests/test_smoke.py etc. import the sibling multicam-occlusion reader, which CI checks out at ../multicam-occlusion but isn't installed locally). This change doesn't touch them.

README gains a short Quickstart snippet for the exporter.