Fix the ScatterD issue in predicated_tile_iterator

[pengpeng-yu]

Summary

This PR fixes the remaining ScatterD pointer-advance issue in PredicatedTileIterator. This change:

• Guards increment_group and increment_cluster in load_with_byte_offset() with !ScatterD
• Updates operator+=() to follow the same pointer-advance rules as operator++()

Fixes #3101.

Validation

I added a minimal local repro bug_repro.zip using:

• CUTLASS GEMM 2.x API
• SIMT core
• gather A + GEMM + scatter D
• int8 x int8 -> int32

Before the fix:

num_mismatch: 8

After the fix:

num_mismatch: 0