fix: align channel loss with sequence parallel labels

[taking-lying-flat]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Fixes #9973.

When sequence parallelism is enabled, per_token_loss_func_sp gathers the per-token loss into the full-sequence layout, while channel-loss metrics still use labels from the local sequence shard. This causes an IndexError and applies the causal label shift twice.

This PR:

• preserves the gathered labels returned by GatherLoss;
• removes SP padding from the gathered loss and labels using the same mask;
• reuses the already-shifted SP labels for channel masking;
• flattens the aligned full-sequence loss and mask before channel slicing;
• keeps packing/padding-free channel boundaries aligned with the full-sequence cu_seqlens;
• flattens the gathered loss before applying per-token loss_scale, covering batch sizes greater than one when no SP padding is needed.