feat: add Metal-backed indexing and search scoring

[aaf2tbz]

Summary

• add centralized wgpu device reporting and lazy cached GPU context
• retain a versioned compact term representation for GPU scoring
• add hybrid Metal term matching for large search batches with CPU fallback
• add configurable GPU thresholds, parity tests, timing/memory instrumentation,
  and exhaustive search benchmarks
• validate native Apple Silicon Metal execution and CPU parity in macOS CI
• document macOS 11+ Metal support and release deployment target

Validation

• cargo fmt --all -- --check
• cargo test --workspace --features gpu
• cargo test -p graphiq-core --no-default-features
• cargo check -p graphiq-bench --no-default-features
• actionlint .github/workflows/*.yml
• local Apple M4 Metal smoke, CPU fallback, and ranked-result parity

The representative Apple M4 exhaustive benchmark (2,119 symbols, 50 uncached
iterations) measured 18,076us median with Metal versus 21,111us CPU-only, with
166 GPU dispatches. The default threshold remains conservative so normal
small interactive searches stay on CPU.