Add Metal GPU backend with MNIST training (98.2% accuracy)

[alok]

Summary

Metal GPU backend for accelerated ML on Apple Silicon.

Features

• GPU-resident buffers with type-safe CpuBuffer/GpuBuffer API
• Optimized GEMM kernels (simdgroup tiling, double-buffered)
• Fused ML ops: biasRelu, biasGelu, biasAdd, softmax, layerNorm
• Flash attention kernels (causal and non-causal)
• Conv2D/MaxPool2D/AvgPool2D for CNN inference
• Mini-batch training with GPU buffer slicing
• Command buffer batching for reduced dispatch overhead

MNIST Results

• 98.2% accuracy on full 60k training set
• 10 epochs, 256-sample mini-batches
• ~230ms per epoch (234 batches)

Files

• Metal/kmeans.metal - Metal shader kernels
• Metal/metal_backend.mm - C++ dispatch layer
• SciLean/FFI/Metal.lean - Lean bindings
• examples/GpuMNIST.lean - Training example