Fix deep-gemm alignment : per-group alignment -> output alignment

[haok1402]

Drop redundant 1024-row over-rounding; per-group 128-alignment already covers FP8 grouped-GEMM tile requirements.

[gemini-code-assist]

Code Review

This pull request simplifies token slicing and removes redundant truncation logic across the codebase. In gpt_oss.py, the manual truncation of input tensors based on group sizes has been removed. In token_scatter.py, the logic for rounding the output token slice to _GEMM_ALLOC_ALIGNMENT was replaced with a direct slice to actual_M, as the group sizes are already guaranteed to be correctly aligned. This adjustment ensures compatibility with DeepGEMM's FP8 kernels, which require the data row count to exactly match the sum of group sizes. I have no feedback to provide as no review comments were submitted.