auto-round

78e057de - feat: Enhance FP8 tensor handling and add cross-shard global scale computation

feat: Enhance FP8 tensor handling and add cross-shard global scale computation - Added `_compute_cross_shard_fused_global_scales` function to compute shared NVFP4 E5M3 global scales for fused groups across different shards. - Improved `_dequant_fp8_tensors` to skip unnecessary hydration when `hydrate_only` is set. - Introduced `_validate_safetensors_file` to check the integrity of safetensors files. - Expanded unit tests to cover new functionality, including streaming scenarios and hydration behavior for FP8 weights. - Ensured that corrupted cached shards are replaced during the prefetch process. Signed-off-by: Xin He <xin3.he@intel.com>