jgrusewski
53382ad692
perf(cuda): unify dual-stream to single stream + event-based sync_all
gpu_backtest_evaluator: removed env_stream, env_event, main_event.
Sequential single-stream loop eliminates 4 event sync calls per step.
H100 forward pass occupies 100+ SMs — env_step overlap was not profitable.
cuda_streams: sync_all() reduced from N blocking cuStreamSynchronize
calls to 1 via cuEventRecord fan-in on stream[0]. 118 tests pass.
Estimated 15-20% walltime reduction in backtest evaluation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>