GemmWithKReduction parallel split-K: workspace initialization is skipped, partials land in the user output buffer

Description

Since CUTLASS 3.2, GemmUniversalBase::initialize() only calls into Params::init_workspace when the mode is plain GEMM:

// include/cutlass/gemm/device/gemm_universal_base.h:415-420
// Assign and prepare workspace memory
if (args.mode == GemmUniversalMode::kGemm) {
  return params_.init_workspace(workspace, stream);
}
return Status::kSuccess;

GemmWithKReduction overrides Params::init_workspace to do its work exactly in the other mode:

// include/cutlass/gemm/kernel/gemm_with_k_reduction.h (init_workspace)
if (this->mode == GemmUniversalMode::kGemmSplitKParallel) {
  ptr_D = workspace;
  ptr_gemm_k_reduction = static_cast<uint8_t *>(workspace)
           + sizeof(ElementC) * size_t(this->batch_stride_D) * size_t(this->grid_tiled_shape.k());
  return Status::kSuccess;
}

The override can therefore never run: in kGemmSplitKParallel mode the base class skips the call entirely. Consequences for device::GemmWithKReduction with --parallel-split-k (example 23):

1. The mainloop writes unreduced per-slice partial products through ptr_D, which still points at the user's D buffer; with a grid whose K extent exceeds the user D allocation this writes out of bounds.
2. The separate reduction launch reads ptr_gemm_k_reduction from a never-initialized workspace.

get_workspace_size does reserve the right size in that mode, so the buffer exists and initialize() returns success; only the redirect is skipped.

Verified statically by reading both functions; additionally exercised on hardware via a probe subclass of the protected base (example 23 flow with --parallel-split-k): after a successful initialize(), params_.ptr_D still equals the user pointer while the reported workspace size is nonzero.

Suggested fix

Widen the gate in gemm_universal_base.h so the virtual init_workspace runs for every mode (the kernels' own overrides already no-op where nothing is needed), or special-case kGemmSplitKParallel alongside kGemm.

Start with include/cutlass/gemm/device/gemm_universal_base.h around the workspace initialization gate, then compare it with Params::init_workspace in include/cutlass/gemm/kernel/gemm_with_k_reduction.h. Exercise the example 23 --parallel-split-k flow or the reported probe and verify that initialization redirects partial output into workspace storage and the reduction reads initialized workspace data.