arch(metal): organize Metal shader library as standalone module

[ohdearquant]

Context

All Metal shaders currently live inline in crates/inference/src/forward/metal_qwen35.rs as Rust string constants (~14K lines, growing). This worked for bootstrapping but is hitting limits:

• No MSL syntax highlighting or linting — shaders are opaque strings to the Rust compiler
• No reuse across models — Qwen3.5-specific but GEMV/softmax/RMSNorm are generic
• Hard to diff — shader changes appear as string diffs, not structured code changes
• Can't benchmark shaders in isolation — coupled to the full forward pass
• Can't study/adapt MLX techniques easily — no clean place to put reference implementations

Proposal

Extract Metal shaders into a dedicated module with .metal source files:

crates/inference/src/metal/
  shaders/
    gemv_q8.metal          # Q8_0 quantized matrix-vector multiply
    gemv_q4.metal          # Q4_0 quantized GEMV
    attention.metal         # decode attention (flash GQA)
    softmax.metal           # online softmax
    norm.metal              # RMSNorm, LayerNorm
    elementwise.metal       # SiLU, residual add
    lm_head.metal           # logit projection + optional argmax
  include/
    common.metal            # shared types, dequant helpers, SIMD-group utils
    quantize.metal          # Q8/Q4 block format definitions
  mod.rs                    # compile-time include_str! or build.rs metal-lib compilation
  dispatch.rs               # kernel launch wrappers (buffer binding, grid sizing)
  pipeline.rs               # MTLComputePipelineState caching

Benefits

• .metal files get IDE syntax highlighting + Metal validation (xcrun metal -c)
• Each kernel is individually reviewable (PR diffs show actual MSL, not string changes)
• Generic kernels (GEMV, softmax) are reusable across model architectures
• Can add a metal-shader-bench binary that benchmarks individual kernels
• MLX kernel adaptations (issue perf(metal): study MLX Metal kernels for GEMV/attention optimization #85) have a clean landing zone
• Build-time xcrun metal compilation catches shader errors before runtime

Migration plan

1. Extract existing inline MSL strings into .metal files (no behavior change)
2. Replace include_str! with metal-rs pipeline compilation
3. Add xcrun metal -c validation to CI (macOS runner only)
4. Refactor metal_qwen35.rs to use the shader library via dispatch.rs

Non-goals

• Not adding new kernels in this issue — just restructuring existing ones
• Not changing the Metal backend API surface
• Not supporting non-Apple GPUs (WGPU backend is separate)

Priority

P2 — prerequisite for sustainable Metal kernel development (issue #85) but not blocking current work.

References

• Issue perf(metal): study MLX Metal kernels for GEMV/attention optimization #85: MLX kernel study
• ADR-025: GPU backend design
• Current Metal code: crates/inference/src/forward/metal_qwen35.rs (~14K lines)