[Metal:Feature] Full W2/W3 decode GEMV + conv1x1 support for Metal ba…

[v0jiuqi]

…ckend

• Deferred dequant + pre-scaling for W2/W3 g8 GEMV
• Extend 2sg decode GEMV + fusions to W_QUANT_2/3
• Extend g16 lm_head GEMV to W_QUANT_2/3
• W_QUANT_2/3 in MetalSharedGather (tied lm_head)
• Complete W2/W3 conv1x1 support
• Vectorize W2 dequant loads, ushort W3 GEMV tiles, size-gate fused W3 GEMM
• Fix W2/3 conv1x1 misrouting on non-simdgroup-matrix GPUs
• Fix W_QUANT_2/3 decode GEMV shadowed by W4 macro alias
• Fix 2/3-bit weight disk embedding decode

Link: https://code.alibaba-inc.com/AliNN/AliNNPrivate/codereview/28994941
GitOrigin-RevId: 9eb095621ebe5f22bdb6cabb3f0e8578ce27c972

Description

Module

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}