Use yaml as the single source of truth for gpu_roce parameters

[melody-ren]

Summary

Make YAML the single source of truth for GPU RoCE and HSB/FPGA transport configuration.

Previously, the same settings could reach the decoding server through YAML, command-line arguments, or QEC_DEVICE_GRAPH_* environment variables. This created ambiguous precedence and made the effective runtime configuration difficult to reproduce.

What changed

• Require YAML transport configuration for:
  • all gpu_roce dispatch paths;
  • all device_graph dispatch paths.
• Remove GPU RoCE and device-graph configuration through environment variables.
• Reject conflicting transport arguments, including --gpu; decoder.cuda_device_id is now authoritative.
• Pass the resolved YAML configuration directly into the device-graph transceiver.
• Update the HSB FPGA demo and test helpers to:
  • copy the source configuration into a temporary launch YAML;
  • add runtime-discovered endpoint details to that YAML;
  • leave the source YAML unchanged;
  • launch the server using only the resulting YAML configuration.
• Use the same 128-byte-aligned GPU RoCE page stride for allocation validation, playback, and the provider.
• Parse sparse decoder matrices through CUDA-QX's canonical YAML parser instead of manually scraping YAML syntax. This supports both flow-style and block-style YAML produced by PyYAML.

Host-only, non-GPU-RoCE configurations retain the existing CLI fallback when no provider is specified in YAML.

Testing

Local container verification:

• Built the affected realtime configuration and YAML test targets.
• Passed:
  • DecoderYAMLTest.RejectsParserErrors
  • DecoderYAMLTest.AcceptsBlockStyleSparseMatrices
  • DecoderYAMLTest.TransportSectionAndMixedDispatch
• Ran formatting and git diff --check.

GB200 hardware verification using the private NV-QLDPC plugin and proprietary cudevice archive:

• extra/hsb-fpga-server/gpu_roce — PASS
• extra/gpu-roce-qldpc-bridge — PASS
• examples/fpga/device-graph/nv-qldpc-decoder — PASS

[melody-ren]

Verified on GB200 using Chuck's handy CI script