T4a deletes the keep-quant provider gates that #2782 landed, dropping coverage from ten encodings to three

[localai-org-maint-bot]

Row: GFX1100-TG200

#2782 lands the kROCM keep-quant GEMM providers and gates them with
tests/vt/test_rocm_quant_dot.cpp: a ten-row encoding table
(IQ2_XXS, IQ3_XXS, IQ2_S, IQ1_S, IQ1_XXXS, Q2_K, Q3_K, Q4_K,
Q5_K, Q6_K) and three cases over it.

51f5222dc in #2790 (T4a) rewrites that file, 497 insertions(+), 231 deletions(-), and the rewrite removes all three:

• ROCm keep-quant GEMM == CPU reference and f64 dequant (Q8_K family)
• ROCm grouped keep-quant GEMM == CPU grouped golden and it WRITES the output
• ROCm keep-quant registers the native kROCM providers

The table goes with them, replaced by kKQuantCases with Q4_K, Q5_K and
Q6_K. The two pull request bodies record the drop without naming it: #2782
says "132,094 assertions green across all ten encodings", #2790 says
"6/6 cases, 719 assertions".

The providers themselves stay. IsRocmKeepQuantSupported still answers yes for
all ten, and rocm_quant_dot.hip still instantiates LaunchGemm and
LaunchGroupedGemm for IQ2_XXS, IQ3_XXS, IQ2_S, IQ1_S, IQ1_XXXS,
Q2_K and Q3_K. After T4a nothing exercises any of them, and nothing
exercises the grouped provider or the registration at all. That is roughly a
thousand lines of kernel with no gate, on the pull request whose own purpose was
to gate them.

Found while gating the external-contributor ROCm stack. Repaired in the same
flow on the landed merge rather than on #2790's branch: #2790 sits under
nineteen later branches that all edit this file, so a commit there costs
nineteen conflict resolutions and this costs none.

The restoration is additive: WeightCase regains its nmse_ref_max override,
kMaxNmseErr comes back, the ten-row table returns as kCases beside T4a's
narrower kKQuantCases, and the three cases return on the
vt::rocm::DeviceAvailable() guard the file now uses. T4a's own cases are
untouched.