[megatron] fix: allow sequence-mean loss with per-token normalization - #7915

What does this PR do?

Fixes #7914 by removing the seq-mean-token-mean rejection in the Megatron per-token loss callback. The guard rejects CP=1 when calculate_per_token_loss=True, and also blocks static-CP THD training even though the forward path reconstructs full-sequence outputs before PPO loss computation.

Checklist Before Starting

• Checked PRs referencing #7914, sequence-mean PRs, and static-CP PRs. No open PR addressing this guard was found; related loss-masking and algorithm changes do not remove this restriction.
• PR title follows the required format.

API and Usage Example

No API changes. loss_agg_mode=seq-mean-token-mean no longer triggers this rejection when calculate_per_token_loss=True.

Design & Code Changes

Deletes only the aggregation-mode guard and its explanatory comment from postprocess_micro_batch_func (13 lines). The existing normalization logic and other validation remain intact.

AI assistance (OpenAI Codex) was used for analysis, implementation, checks, and PR preparation under the issue author's direction.

Checklist Before Submitting

• Read the contribution guide.
• Ran pre-commit hooks for the changed file; the config hook passed after selecting the development Python environment as noted above.
• Documentation changes: none for this guard removal; rationale is in [megatron] seq-mean-token-mean guard rejects CP=1 and appears overly restrictive for reconstructed THD outputs #7914.
• New unit/end-to-end tests: none requested; validation and limitations are listed above.
• CI request in Slack/Feishu: not sent.