feat(nemotron_v3): run MTP heads under eval for validation acceptance metrics (#2802)

* feat(nemotron_v3): run MTP heads under eval for validation acceptance metrics

Add a `compute_mtp_in_eval` flag (default False) to NemotronHForCausalLM so the
MTP heads can run during validation to measure per-head token acceptance, while
keeping the cached generation/decoding path free of any MTP cost (gated on
`not use_cache`). The training harness toggles it around the validation forward
only.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: SlyneD <slyned@nvidia.com>

* test(nemotron_v3): cover MTP configuration and checkpoint views

Signed-off-by: SlyneD <slyned@nvidia.com>

---------

Signed-off-by: SlyneD <slyned@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>