Worker A merged: IQ2_S Metal fix (constant memory → device buffer)

Critical fix: iq2s_grid[1024] (8KB) in Metal constant memory
exceeded Apple Silicon per-function limit. Moved to device buffer.

Metal MoE now functional (no hang) but per-phase dispatch is slower
than CPU fused IQ2 dot. Single command buffer hangs (phase data
dependency issue). Metal MoE disabled; CPU path at 1.5-3.7 tok/s.

IQ2_S standalone test passes. IQ2_XXS matmul test passes.
All fixes preserved for future single-cmdBuf investigation.

32/32 tests, "Paris" ✓ on 35B.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

.claude/worktrees/agent-afba980a

@@ -1 +1 @@
-Subproject commit 3a53020d65d75537bd22ee8e818c0d96b8428632
+Subproject commit d416018be33659d81fbd129efd78428ec517ddcc