MToon: specialize shader pipelines via [[function_constant]] feature flags

## Source
Romero perf review, item 4.

## Problem
The MToon shader is an ubershader — branches at runtime on uniform flags (matcap, rim, alpha-cutout, shade-shift texture, …). Apple Silicon handles uniform branches well, but the compiler still spills registers defensively for the worst-case path.

## Proposed change
Use Metal `[[function_constant(n)]]` to compile a dedicated pipeline variant per material's actual feature set. Extend `VRMPipelineCache` key from `(blendMode, …)` to `(vertexDescriptor, featureMask, blendMode)`. Variants compile lazily as new materials are seen; the cache amortises the cost.

## Affected
- `Sources/VRMMetalKit/Renderer/VRMPipelineCache.swift`
- `Sources/VRMMetalKit/Shaders/MToon.metal`
- `Sources/VRMMetalKit/Renderer/VRMRenderer.swift` (variant selection at bind time)

## Related
- #155 (MToon shader maintenance)
- #125 (per-light dedup — should land first or be coordinated)
- #121 (shared shader header — same area)
- #156 (measure)

Start with Sources/VRMMetalKit/Renderer/VRMPipelineCache.swift and Sources/VRMMetalKit/Shaders/MToon.metal, then trace variant selection in Sources/VRMMetalKit/Renderer/VRMRenderer.swift. Determine how feature flags map to function constants and how the cache key and lazy compilation interact with material binding. Done means feature-specific MToon pipeline variants are selected and cached correctly; coordinate with #125 and review related work in #155 and #121.

Written by the indexing model from the issue text.