feat(pp): support FusedLinearCrossEntropy under pipeline parallelism

[beccohov]

What does this PR do ?

Makes FusedLinearCrossEntropy actually run under pipeline parallelism instead of
silently falling back to MaskedCrossEntropy on every PP run (#2681).

Changelog

• recipes/llm/train_ft.py: run the loss-capability check on the real last-stage
  module, not the AutoPipeline wrapper (which defines no forward, so the check
  always failed and silently swapped FusedLinearCrossEntropy for
  MaskedCrossEntropy). Extracted into _maybe_downgrade_loss_fn.
• recipes/llm/train_ft.py: gate fused CE under PP to the generic patched forward —
  MoE/custom-forward models (e.g. deepseekv3) fall back to MaskedCrossEntropy with a
  clear warning instead of crashing; flag the last stage to emit hidden states when
  fused CE is active.
• distributed/pipelining/hf_utils.py: the last stage skips the lm_head projection
  and returns hidden states when flagged; mark the model fused-CE-capable only when the
  generic forward is installed.
• distributed/pipelining/functional.py: size the last-stage output as
  [mb, seq, hidden] instead of [mb, seq, vocab] on the fused path.
• components/loss/mtp.py: PipelineCausalLMLoss treats a bare last-stage tensor as
  hidden states for FusedLinearCrossEntropy (running the fused lm_head+CE) and raises
  a clear error for the unsupported MTP-tuple + fused-CE combination.
• Tests: forward emits hidden states; stage output shape switches to hidden; pipeline
  loss routes the bare tensor to hidden states; capability/gate truth table.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

Additional Information

• Related to FusedLinearCrossEntropy silently no-ops under pipeline parallelism (and doesn't function under PP) #2681
• Scope: generic HF causal-LM PP path (covers the Llama-3.3-70B repro). MoE/MTP
  custom-forward models (deepseekv3, nemotron_v3, step3p7) are gated to
  MaskedCrossEntropy and left as a follow-up.
• Validation: PP=4 × TP=2 tiny Llama — fused CE trains without the fallback
  warning and produces bit-identical loss/grad-norm to MaskedCrossEntropy across steps
  (confirms correctness under TP-sharded vocab).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[beccohov]

Hey @akoumpa,
as promised -- here is the PR!