mindtorch_v2/mps: real Metal GPU compute backend for 30+ ops

[lvyufeng]

Summary

• Replace NumPy CPU fallbacks with actual Metal GPU compute kernels dispatched via MTLComputeCommandEncoder
• Add 105 MSL (Metal Shading Language) kernels covering element-wise binary (9), unary (16+2 activations), reduction (6), softmax, fill, copy, and in-place (8) operations — all with f32/f16 variants
• Add MetalKernelDispatcher engine: lazy shader compilation, pipeline state caching, and thread dispatch helpers (PyObjC + ctypes dual-path, no new dependencies)
• Implement GPU matmul via Apple's MPSMatrixMultiplication kernel — 4.3x faster than Accelerate BLAS at 1024×1024
• All f64 and non-contiguous tensors automatically fall back to existing CPU/Accelerate path

New files

File	Purpose
metal_shaders.py	All MSL kernel source strings (generated from templates)
metal_compute.py	GPU dispatch engine: compilation, pipeline cache, encode+commit

Modified files

File	Changes
runtime.py	+compile_library, make_compute_pipeline, get_compute_encoder (PyObjC + ctypes)
mps_kernels.py	GPU matmul via MPSMatrixMultiplication (PyObjC + ctypes fallback)
ops.py	30+ ops upgraded: GPU-first dispatch with CPU fallback

Test plan

• All 27 correctness checks pass (binary, unary, reduction, matmul, activations, in-place, softmax)
• torch.allclose validation against CPU reference for every op
• MatMul 1024×1024 benchmark: GPU 3.71ms vs CPU 16.10ms (4.3x speedup)
• f64 tensors correctly fall back to CPU path
• Non-contiguous tensors correctly fall back to CPU path
• Run full model test suite (BERT, GPT-2) on MPS device

🤖 Generated with Claude Code