Add Dockerized basic Holoscan ingest pipeline

[dleshchev]

Summary

• Add a Docker-built Holoscan v4.4.0 CUDA 13 pipeline image.
• Add native C++/CUDA Holoscan operators:
  • ZmqRxOp: ZMQ receive plus fast CBOR bslz4 extraction
  • DecompressGpuOp: GPU bslz4 path using nvCOMP LZ4 plus CUDA bit-unshuffle
• Add the Python-composed pipeline:
  ZmqRxOp -> DecompressGpuOp -> StatsSinkOp
• Emit decompressed frames as CUDA tensors and consume them from Python with PyTorch DLPack.
• Keep the existing Python pipeline in place while the Holoscan path is validated.

Validation

• python3 -m py_compile holoscan_pipeline/python/dectris_holoscan/*.py
• docker compose -f docker/compose.holoscan.yaml config
• git diff --check
• docker build -f docker/Dockerfile.holoscan -t dectris-holoscan-basic .
• GPU container import smoke test for dectris_holoscan_ops
• Full end-to-end 1,000,000 image run:
  • simulator source rate: 139,136.5 frames/s
  • pipeline processed 1,000,000 frames in 1000 batches
  • final sink rate: 139,608.3 fps
  • output tensor: shape=(1000, 96, 96), dtype=torch.uint16, device=cuda:0
  • pipeline errors: 0

Current Limitations

• Fixed initial detector assumptions: 96x96 uint16 bslz4 frames.
• nvCOMP is inherited from the Holoscan base image (libnvcomp5-cuda-13 at 5.0.0.6-1 in the validated image).
• CompressedBatch payloads point into a three-slot pinned host pool owned by ZmqRxOp; the basic pipeline relies on downstream consumption before slot reuse.

Follow-ups

• Add focused correctness tests against known frames.
• Generalize detector shape, dtype, and compression settings beyond the initial binned ARINA path.
• Replace the implicit pinned-pool producer/consumer contract with explicit slot ownership or scheduler backpressure.
• Consider preallocating nvCOMP per-batch workspace once the target batch/shape set is finalized.