[GEMM] Fix sparse split-K metadata offsets (issue #3533)

[XFDG]

Summary

Correct the packed metadata E column offset for K slices after the first in three sparse GEMM kernel bodies:

• SparseGemm
• SparseGemmWithAbsmax
• SparseGemmWithEpilogueVisitor

Each offset now divides by kElementsPerElementE, matching the metadata extent and GemmSparseUniversal.

Rationale

The A and B slice offsets are expressed in their respective logical element coordinates. E is different: each stored metadata element describes kElementsPerElementE groups of sparse A values, and its iterator extent already divides by both kSparse and kElementsPerElementE.

The old E offset divided only by kSparse. For split-K slices greater than zero, this started metadata iteration beyond the packed E slice. Predication then suppressed valid metadata loads and the slice accumulated against invalid zero-filled metadata.

Validation

Used the non-universal SparseGemm path from example 15 with serial split-K enabled. Built with CUDA 13.1 and -arch=sm_100a and ran on an NVIDIA B200.

Control:

• split-K=1 passes the complete example reference comparison on current main

Before:

• split-K=2 fails the same end-to-end reference comparison

After:

• split-K=2 passes the end-to-end reference comparison
• all three changed headers compile in the same CUDA translation unit
• git diff --check passes

This path is distinct from the existing GemmSparseUniversal split-K test coverage.

Fixes #3533