[GEMM] Fix EllGemm column-major split-K workspace sizing (issue #3539)

[XFDG]

Summary

Make the ColumnMajor EllGemm specialization size its serial split-K semaphore workspace from the same transposed problem shape used by initialize().

Add a regression test for both rectangular orientations, covering the previous under-allocation and over-allocation cases.

Rationale

ColumnMajor output is implemented by exchanging M/N and forwarding to a RowMajor underlying operator. initialize() correctly computes its launch grid from (n, m, k), but get_workspace_size() used the original (m, n, k).

For rectangular problems, the two shapes can produce different tile counts. With m=264, n=136, ell_blocksize=32, and four split-K slices, workspace sizing returned 15 semaphore slots while the initialized grid contains 18 output tiles. This left part of the semaphore array outside the allocation/initialization contract.

Using the transposed problem shape in get_workspace happens to match initialize() and the underlying operator for every aspect ratio.

Validation

Compiled the same standalone reproducer before and after this change with CUDA 13.1 and -arch=sm_100a, then ran it on an NVIDIA B200.

Before:

• initialized launch grid: 2 x 9 tiles
• get_workspace_size reported 60 bytes instead of 72
• initialize() returned success but cudaMemsetAsync zeroed only 60 of the required 72 bytes

After:

• get_workspace_size reports 72 bytes
• initialize() zeroes all 72 bytes
• the opposite 136 x 264 aspect ratio is also pinned by the new unit test at 60 bytes
• the new unit-test translation unit compiles successfully for sm_100a
• git diff --check passes

Fixes #3539