Add TensorRT deployment and ResNet-50 detector support

[haruto2002]

Summary

Two additions from post-challenge work, one commit each:

1. TensorRT deployment (trt/) — a reproducible toolchain that takes a
   Sparse4D checkpoint to a real-time FP16 engine.
2. ResNet-50 detector support — a ready-to-use config for the public NGC
   sparse4d_warehouse_v2.2_r50 checkpoint, quality-weighted detection
   scoring, and documentation of the findings behind both.

Neither changes existing behavior: the paper's published numbers remain
exactly reproducible with the documented settings.

TensorRT deployment

• Full chain .pth → ONNX → engine → inference → GT check, run entirely
  inside the vss-rt-cv container (host needs only docker + GPU driver).
• ONNX export with automatic TAO spec repair, trtexec build wrapper
  (FP16 + MSDA plugin, dynamic camera count), engine inference with
  externally driven instance-bank recurrence, and a hybrid runner pairing
  the TRT detector with the PyTorch identity layer.
• Includes a defense against an FP16 recurrent-cache NaN collapse
  (nan_to_num + clamp at the feedback boundary; reproducible with
  --no-clamp, also reported to NVIDIA).
• 4.6× speedup at 16 cameras — 32.5 ms/step (30.8 FPS, real time) vs
  149.3 ms in PyTorch — for a 0.4–1.8 HOTA cost. Verified on all three
  validation scenes × 9,000 frames against the official 3D-HOTA protocol
  (trt/RESULTS.md).

ResNet-50 support & quality scoring

• src/sparseqid/configs/sparse4d_rn50_v2.2: architecture config + anchors
  for the public rn50 checkpoint (use via --config-dir).
• sqid infer --quality-score: scores detections as
  sigmoid(cls) × sigmoid(centerness) (the TAO decoder convention) instead
  of the class score alone. Opt-in; default behavior is unchanged.
• docs/EXPERIMENTS.md records the post-challenge ablation study: the rn50
  detector is far more robust under domain shift (val GT-weighted HOTA
  25.64 → 35.01), float32 inference is worth several HOTA points over
  --amp bfloat16 (+4.6 on the hidden test set), and quality scoring adds
  ~+1. The full ladder reaches 38.69 on val and 55.9 on the hidden test set
  (submitted system: 38.01). docs/REPRODUCE.md documents the
  precision/reproducibility trade-off of --amp.

Testing

• uv run pytest: 14 passed, 1 skipped (unchanged).
• TRT chain verified end to end on Warehouse_020–022 (accuracy + timing in
  trt/RESULTS.md).
• --quality-score output verified byte-identical to the externally
  validated driver used in the experiments.