Sm90VisitorImplBase fixed-size specializations advance workspace by raw sizes while get_workspace_size/initialize_workspace round to 16

Description

In include/cutlass/epilogue/fusion/sm90_visitor_tma_warpspecialized.hpp, the hand-written Sm90VisitorImplBase specializations for 2, 3, and 4 ops advance per-op workspace offsets by the raw Op::get_workspace_size(...) value in to_underlying_arguments, while their own get_workspace_size and initialize_workspace round every op's size up to MinWorkspaceAlignment (16) before advancing. The generic variadic base rounds consistently in all three functions; only these specializations diverge.

• 2-op specialization: lines 995-997 (op_1_workspace = op_0_workspace + op_0_workspace_size, no rounding), vs rounding in 1013-1022 and 1030-1041
• 3-op specialization: lines 1093-1097
• 4-op specialization: lines 1210-1216

Consequence

If an earlier op's workspace size is not a multiple of 16 and any later op has nonzero workspace, Params produced by to_underlying_arguments point that later op at [ws+raw, ws+raw+size) while initialize_workspace initialized [ws+round16(raw), ...). On device the later op then reads/writes a region that was never initialized and partially overlaps the previous op's workspace.

Today only Sm90RowReduction / Sm90ColReduction allocate workspace among stock ops, and their trailing tile-counter block (ceil_div(N, tile_N) * sizeof(int), line 1156 / ceil_div(M, tile_M) * sizeof(int), line 1745) is not always a 16-multiple: e.g. N=600 with tile_N=128 gives ceil_div=5 -> 20 bytes. So the divergence is reachable whenever such a reduction sits at position < last in a <=4-op EVT tree followed by another workspace consumer. I have not constructed an end-to-end failing fusion; this is a static analysis of the offset bookkeeping, filed because the inconsistency is objective.

Suggested fix

Round each op's size to MinWorkspaceAlignment when advancing offsets in the three to_underlying_arguments overloads, matching the variadic base (lines 504-507). Alternatively delegate the fixed-size specializations to the same shared helper so the two paths cannot drift again.