test(unit): ~26 hardware-gated tests report as passed instead of skipped

[jamesburton]

Observation

Found while addressing review feedback on #315, where one test used this shape:

if (!Avx2.IsSupported) return;   // reports as PASSED on hardware without AVX2

A test that returns early without asserting is reported as passed, not skipped. On a CI agent or
developer machine lacking the required ISA, the suite goes green while the coverage silently
evaporates — which is the failure mode a test suite exists to prevent.

grep puts roughly 26 further occurrences across tests/DotLLM.Tests.Unit, gated on things like
Avx2.IsSupported, Avx512F.IsSupported, CUDA availability and Vulkan device presence.

Why it matters here specifically

This repo's correctness story leans on parity tests — SIMD kernels verified against a scalar
reference, backends verified against CPU. Those are exactly the tests that carry hardware gates, so
the pattern concentrates in the highest-value tests rather than the incidental ones.

It also interacts badly with heterogeneous hardware: a contributor on a machine without AVX-512 sees
green, pushes, and only discovers the gap if someone else happens to run it on a machine that has it.

The convention already exists

tests/DotLLM.Tests.Unit already references Xunit.SkippableFact and uses it in the CUDA tests:

[SkippableFact]
public void Something()
{
    Skip.IfNot(Avx2.IsSupported, "Requires AVX2.");
    ...
}

So this is applying an established convention consistently, not introducing one.

Proposed change

1. Convert hardware-gated early returns to [SkippableFact]/[SkippableTheory] + Skip.IfNot.
2. Where a test runs BOTH a scalar and a vector path, gate only the vector half — skipping the whole
   test would discard the scalar coverage, which does run everywhere. (This distinction was applied
   in kernels(cpu)(matmul): F32 outer-product tiled GEMM kernel for prefill #315 and is the reason that PR did not convert two of its tests.)
3. Leave a short comment at any site where the early return is genuinely intentional.

Acceptance criteria

• No test in tests/DotLLM.Tests.Unit returns early on a hardware capability check without
  either skipping or retaining an assertion that runs everywhere.
• Skipped counts appear in the run summary on hardware lacking each capability, so the gap is
  visible rather than silent.
• Tests exercising both scalar and vector paths still assert the scalar path when the vector
  path is unavailable.
• No change to which assertions run on fully-capable hardware — the pass count there is
  unchanged.

References

• kernels(cpu)(matmul): F32 outer-product tiled GEMM kernel for prefill #315 — where the pattern was found; that PR converts the one occurrence in its own diff
• tests/DotLLM.Tests.Unit/Cuda/ — existing SkippableFact usage to follow