A vLLM ROCm fix found by kernel-timeline fingerprinting, after two source-reading guesses were wrong.

PR #58566 moves a contiguous copy in rocm_unquantized_gemm_impl into the only two branches that use it. Shapes passing the broad skinny-GEMM gate but matching neither branch were paying for a copy that got discarded. On Kimi-K3 FP4 across 8x MI355X, per-step copies dropped from 73 to 4, saving 0.339ms per decode step, almost all kernel-launch overhead. The author notes both earlier hypotheses from reading source were wrong, and only the profile found it. Good reminder for the "let the agent read the code and reason about perf" workflow.