ROCm skinny GEMM architecture eligibility caches the first device

[VikashLoomba]

Problem

PR #506 adds the wave32-only ROCm wvSplitK skinny GEMM. Its dispatcher currently implements:

bool SkinnyGemmArchOk(int device_index) {
  static const std::string arch = vt::rocm::DeviceArchName(device_index);
  ...
}

The function accepts a device index but caches only the first device's architecture. In a heterogeneous process, a first call on gfx11/gfx12 poisons later calls on gfx9 as eligible. The adjacent source comment says the unported wave64 arm would then produce silently wrong sums.

A fresh integration reviewer found this while rebasing #506 onto main's dual-slot GetBlas/device-hop seam. The existing focused test uses only device index 0, so it cannot detect the defect.

Required fix

• Cache/evaluate eligibility by device, without adding a per-token HIP property probe or a hot-path global lock.
• Add a RED-first device-hop test using a deterministic architecture resolver/cache seam, because this host's four GPUs are all gfx1100.
• Preserve the wave32-only gfx11/gfx12 policy, unknown-architecture refusal, main's dual-slot GetBlas behavior, and all existing wvSplitK guards.
• Mutation-prove that collapsing eligibility back to the first device makes the new test fail.

Ownership

Owning row: BACKEND-ROCM. Fix in the existing PR #506 flow before its rebased head is pushed.

[VikashLoomba]

Fixed in reviewed PR #506 at immutable head a759a0746a09342febf43f0624f21796d9410d74. The repair keys architecture eligibility by resolver and device; RED/GREEN mutation evidence and current-main validation are recorded on the PR.