fix(vmi): materialize compact group slots as contiguous

[chenshengxin2026]

Fixes #1377.

Problem

VMILowerUnifiedToLegacy rewrites every continuous pto.vmi.vload of 1/2/4/8
lanes into a pto.vmi.group_slot_load, because that is the only lowering that
can express a bounded short load (pto.vsldb, a 32-byte block under a
PAT_VL{n} mask). The dense path cannot: it needs whole physical chunks or a
provable whole-register over-read, which a !pto.ptr source can never supply.

The side effect is that the value's layout becomes num_groups instead of
contiguous, so histogram and scatter consumers that require a contiguous
layout ask for an ensure_layout conversion that was never registered:

ensure_layout has no registered materialization support:
source/result layouts do not match a supported ensure_layout table row

Fix

Register the relation as a physical invariant rather than as a list of the
group counts the alias happens to emit.

A group packet with num_groups <= slots, lane_stride = 1 and
num_groups <= lanesPerPart puts logical lane i on physical lane i of a
single carrier — exactly what a dense contiguous short vector does
(getVMIPhysicalArity is 1 on both sides; mapLogicalLaneToPhysical agrees
lane for lane). The conversion is therefore a pure register forward, in either
direction, with no pack/zip/shuffle.

• New table key gsFit() in the pattern DSL, used by two symmetric rows in
  kEnsureLayoutPatterns. It subsumes the previous one-group/one-slot rows.
• isVMISingleCarrierGroupSlotAlias (in VMILayoutSupport.h) is shared by the
  table matcher and the VPTO materializer, so the two cannot drift.

This also closes neighbouring cases that a group-count list leaves open:
an explicit vload {group = 3} hits the identical diagnostic today, and so
does the mirror direction, where a dense short vector reaches the compact
group_store alias.

group_slot_load keeps its !pto.ptr source requirement. pto.vsldb is
pointer-only, and PTODSL always hands pto.vmi.vload a pointer (kernels build
UB buffers with pto.castptr(..., pto.ptr(dtype, "ub"))), so nothing in the
DSL path needs a memref source here.

Scatter

Small-VL vscatter is not made to work. It reaches the pre-existing
full-physical-chunk gate and is rejected with a stable
VMI-UNSUPPORTED: pto.vmi.scatter ... value requires full physical chunks
diagnostic, which is resolution path 2 in the issue. Documented in
docs/isa/vmi-isa/07-sfu.md and the PTODSL user guide.

Validation

• Reproduction matrix: 21/21 checks pass (vchist/vdhist VL=1/2/4/8 lower;
  VL=64/128/256 controls unchanged; vscatter VL=1/2/4/8 rejected with the
  documented diagnostic).
• vmi_new lit suite: every test whose RUN lines use pto-test-opt passes.
• New/updated coverage: num_groups = 3, the reverse contiguous -> group
  direction, a 128-bin (Bin_N0-only) accumulator matching the shape reported
  in the issue, and negative cases for num_groups > slots,
  lane_stride = 2, and a group count wider than one carrier.

Note for reviewers

vmi_layout_gate_helper_support_invalid.pto pinned
contiguous -> num_groups = 8, slots = 8 at VL=8 as unsupported. That is
precisely the mirror direction this change enables, so the test was repointed
at num_groups = 64, slots = 8, which stays unsupported and preserves what the
test was checking.

Follow-up

The alias itself still welds a load strategy to a layout attribute:
VMIGroupSlotLoadOp::verify forces a num_groups result layout even when the
value is physically indistinguishable from a contiguous one. Decoupling those
is tracked separately and is out of scope here.