Add M16 hardware-portable ML execution

[alanthssss]

Summary

• add explicit auto, cpu, mps, and cuda selection to every ML command
• prefer CUDA, then Apple MPS, then CPU for auto, while refusing silent fallback for explicit accelerator requests
• record architecture, device identity, Python, PyTorch, and available backends in M6–M14 run metadata
• add accelerator-synchronized BC timing, throughput, and CUDA peak-memory evidence
• make CTDE rollout workers initialize their hardware backend correctly
• add backend selection regression tests and equal English/Chinese M16 documentation
• update README, handoff, product design, and both changelogs with overall project progress

Local evidence

Apple M1 Pro, PyTorch 2.13.0, fixed 20-episode BC dataset, 3 epochs:

Backend	Time	Throughput	Validation	Test
CPU	0.5774 s	138,355 examples/s	50.50%	48.96%
MPS	3.6764 s	21,730 examples/s	49.34%	48.57%

This small workload was about 6.4× faster on CPU, so AWS CUDA is deliberately deferred until profiling shows a suitable workload.

Validation

• 53 core tests passed
• Python compilation passed
• M6 BC+PPO smoke pipeline completed on MPS
• M7 CTDE smoke pipeline completed on MPS
• git diff --check