perf(inference): fused RMSNorm+Residual kernel to reduce dispatch count by 48/token

[ohdearquant]

Problem

Each of the 24 transformer layers has 2 normalization points, each currently requiring 2 Metal dispatches: one for RMSNorm, one for residual addition. That's 24 × 2 × 2 = 96 dispatches just for norm+residual.

Proposal

Fuse RMSNorm + residual add into a single kernel per normalization point:

• Read input once
• Compute variance → normalize → scale → add residual
• Write output once

This cuts 48 dispatches/token (from 304 to ~256), a 16% dispatch reduction.

Evidence

• metal_qwen35.rs line ~8663 (GDN) and ~9043 (GQA): norm and residual dispatched separately
• Metal best practice: minimize dispatch count to reduce command-buffer overhead
• At 304 dispatches/token with ~0.02ms overhead/dispatch, 48 fewer = ~1ms/token potential

Acceptance

• Single kernel handles: output = RMSNorm(x) * weight + residual
• No PPL regression (±0.01 vs reference)
• bench-compare shows measurable intercept improvement
• clippy clean

Priority

P1