fix(vmi): materialize packed group-slot widening

[Zhendong404]

Summary

• materialize compact group_slots widening with vcvt partition streams and register interleaves
• support 2x and 4x storage-width widening for floating-point and integer VMI extension lowering
• derive and validate 256-byte physical carriers per chunk; avoid invalid logical-size VPTO vregs
• add BF16 8/12-group and UI8 widening coverage through public VMI ops

Scope

This PR fixes the lowering required to express the ASC-style group-result conversion/store sequence. It does not rewrite the #1209 reproducer end to end or establish an A5 performance result by itself.

Root cause

A grouped BF16 reduction produces compact group_slots values. Widening those values partitions compact source lanes (EVEN/ODD for 2x, P0..P3 for 4x); the old lowering had no legal materialization that restored their original lane order before a group store.

The current #1209 VMI fixture works around this limitation by widening the full abs_bits vector to f32 and running a second f32 group reduction for scale_out. ASC instead reuses the BF16 grouped maximum directly (vcgmax -> bitcast -> vcvt PART_EVEN -> store). Thus this PR removes the lowering blocker, but the fixture must still be rewritten to consume the original group maximum before the VMI source algorithm aligns with ASC.

Lowering

• 2x: vcvt EVEN, vcvt ODD, then vintlv.low
• 4x: vcvt P0..P3, then three vintlv.low operations

slots=1 remains on the existing sparse path and is deliberately excluded.

Validation

• full local ninja build
• llvm-lit 7 focused VMI tests, all passed
• git diff --check

Relates to #1209.