v0.33.0-rc0

[borg323]

This release includes improvements across the cuda, onnx and demux backends, along with several search, benchmarking, and training-data fixes.

• Several DAG search improvements: concurrent task workers, improved lock contention, garbage collection, more accurate draw score as well as better time keeping.
• Added cuda backend graph support and cutlass fused multihead attention.
• Refactored demux to reduce lock contention and improve batch handling.
• Added support for preferred batch step selection in the cuda, onnx and xla backends.
• Added support for network evaluations per second in uci info, enabled by the --show-eps command line switch or UCI_ShowEPS uci option.
• Use cuda streams if possible in onnx, for more efficient memory transfers.
• Switched onnx-trt to strongly typed graphs to avoid problematic type conversions in TensorRT, added quantization control through the optimize backend option and full bf16 support.
• Added the onnx-coreml backend.
• Also added the onnx-migraphx backend, with fixed batch sizes recommended for best performance.
• Allowed the rescorer to process v7 training data.
• Added tracing support.
• Search avoids allocating extra computations for better backend buffer reuse.
• Improved backendbench statistics and added thread support.
• Added back a threads option to the blas backend.
• The master version of net.proto is now hosted by lc0 and includes support for fp16 and fp32 weights encoding.
• Notable bug fixes:
  • Restored the --history-fill naming.
  • Reverted ponder behavior changes.
  • Fixed castling move display for FRC.
  • Fixed multiplay issues.
  • Fixed a DAG n_in_flight underflow bug.
  • Defaulted cuda backend option fused_mha to false for networks with relu^2 activation.
  • Fixed a segfault in Python bindings for transformer networks when calling Weights.filters().
  • Fixed onnx locking.
• Several build system updates.
• Assorted small fixes and improvements.

---

This discussion was created from the release v0.33.0-rc0.