Add Apple Metal (MPS) and CPU support with auto device selection

[james-see]

Summary

Scorer and benchmark runners now work on Apple Silicon (MPS) and CPU in addition to CUDA, with automatic device selection (CUDA > MPS > CPU).

Changes

• core.py: new `resolve_device()`, `resolve_dtype()`/`dtype_candidates()` (bf16 preferred, fp16/fp32 fallback off CUDA), `synchronize()`, `describe_hardware()`, portable peak-memory helpers. `load_causal_model()` accepts optional `device`/`dtype`; CUDA path stays strict (exactly-one-GPU, bf16-only) so published results remain comparable; actual dtype/device recorded in metadata.
• direct/reranker/shared/serial: `torch.cuda.synchronize` branches replaced with portable `synchronize(device)`.
• CLI + shape777 / shape777_reranker / decision_vs_generation: new `--device`/`--dtype` flags (default `auto`); portable hardware/memory reporting (`peak_cuda_bytes` kept for compat, plus `peak_memory_bytes` + `device`).
• Docs: AGENTS.md one-accelerator policy; benchmarks/README Apple Silicon usage; METHOD.md MPS/CPU non-comparability note.
• tests/test_device.py: 14 tests for resolution priority, dtype policy, sync safety, revision validation.

Validation

• `pytest tests/ -q`: 26 passed (incl. live MPS smoke on ARM64 Mac)
• `(cd results/raw && sha256sum -c SHA256SUMS)`: 15 OK
• `python benchmarks/verify_published.py`: 69 claims, status ok
• No headline claims, phase1-summary.json, or committed raw outputs touched."

[james-see]

@TheoLeeCJ this should be gtg now no conflicts. need to support MPS / mac metal vs. just CUDA

[james-see]

@TheoLeeCJ any movement on this?

[TheoLeeCJ]

Sorry for the wait. Other PRs now cover Apple Silicon and llama.cpp CPU support, but the Torch CPU and device-selection work here remained useful. I ported that focused portion onto current master in 23cf1f3 and preserved attribution to your contribution. Thank you for the work!