Fix grouped GEMM launch code to derive grid size from filtered kargs

[Copilot]

Kernel::MakeKargs(gemm_descs) skips groups where M/N/K==0, so it can return fewer entries than gemm_descs.size(). The grid size was computed via Kernel::GridSize(gemm_descs) which used the unfiltered count, causing the kernel to be launched over more blocks than there are valid kernel arguments.

Changes

• Grid size: Kernel::GridSize(gemm_descs) → dim3(kargs.empty() ? 0 : kargs.back().block_end, 1, 1) — derived from the filtered kargs vector, consistent with how hipMemcpyWithStream and group_count already used kargs.size()

// Before
const dim3 grids = Kernel::GridSize(gemm_descs);  // uses gemm_descs.size() — can exceed kargs

// After
const dim3 grids = dim3(kargs.empty() ? 0 : kargs.back().block_end, 1, 1);  // aligned with MakeKargs output

The hipMemcpyWithStream byte count and group_count kernel argument were already correctly using kargs.size().

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.