[VMI] Support L=8 ui8 to ui16 widen on the fused scale path

[mouliangyu]

Summary

The fused quant + dequant path (per_token_cast_and_cast_back) needs a compact
ui8 -> ui16 widen for an L=8 scale-factor value. When the value is represented as
the VMI group-slot layout used by the reproducer, VMI legalization leaves a residual
pto.vmi.extui instead of lowering it to VPTO.

Reproducer

Reproducer PR: #1015

The PR contains:

• an L=256 widen that currently lowers;
• the failing L=8 VMI fixture;
• a working MI/CCE reference;
• compile logs showing the residual illegal extui.

Pinned repro version: vmi-v0.1.3.

Expected behavior

For the actual layout produced by the fused quant/dequant kernel, VMI ui8 -> ui16
widen should lower to legal VPTO conversion/store operations without a residual VMI
cast. The expected physical layout must match the source value: a group-slot value and
a compact contiguous 8-lane value are different cases and should not be conflated.

Acceptance criteria

• confirm the exact layout of the value at the per_token_cast_and_cast_back_vmi.py
  call site;
• add a minimal PTOAS test for that layout and L=8;
• lower it without residual pto.vmi.extui;
• verify the generated conversion uses a legal mask/part selection and preserves the
  scale values;
• compile the affected fused kernel path without a compose/separate-cast fallback.

Related kernel reference (if accessible):

• https://github.com/deepseek-ai/TileKernels-Nightly/pull/80