feat: add Metal backend for Apple GPU support

[GiggleLiu]

Summary

• Add new tropical-gemm-metal crate for GPU-accelerated tropical GEMM on macOS
• Support MaxPlus, MinPlus, and MaxMul semirings (f32)
• Include argmax tracking and backward pass kernels for neural network training
• Follow the same architecture as the CUDA backend for consistency

Implementation Details

The Metal backend provides:

• MetalContext: Compiles and caches Metal shaders, manages command queue
• GpuMatrix: GPU memory management with row-major ↔ column-major conversion
• MetalKernel trait: Type-safe kernel dispatch for each semiring type
• Global context caching: Avoids shader recompilation overhead

API Example

use tropical_gemm_metal::{tropical_matmul_metal, MetalContext};
use tropical_gemm::types::TropicalMaxPlus;

// One-shot API with cached global context
let c = tropical_matmul_metal::<TropicalMaxPlus<f32>>(&a, m, k, &b, n)?;

// Or with explicit context for repeated operations
let ctx = MetalContext::new()?;
tropical_gemm_metal::<TropicalMaxPlus<f32>>(&ctx, &a_gpu, &b_gpu, &mut c_gpu)?;

Test plan

• All 6 unit tests pass on Apple Silicon
• MaxPlus, MinPlus, MaxMul semirings tested
• Argmax tracking verified for backpropagation
• Larger matrix (64x128x64) test passes

Closes #1

🤖 Generated with Claude Code