Ignore decoder_inputs_embeds when decoder_input_ids are present (fixes #39542)

[JoseAlvarezMedina]

Fixes #39542

What does this PR do?

When using a custom seq2seq model together with PEFT/LoRA, both decoder_input_ids and decoder_inputs_embeds can end up being passed to the underlying decoder. This triggers the internal validation:

ValueError: You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time

This PR adds a small defensive override in Seq2SeqTrainer.compute_loss that drops decoder_inputs_embeds when decoder_input_ids are also present. This keeps backward compatibility and mirrors the user expectation that decoder_input_ids should take precedence.

Implementation details

if "decoder_input_ids" in inputs and "decoder_inputs_embeds" in inputs:
    inputs.pop("decoder_inputs_embeds")
A new test tests/trainer_seq2seq_test.py builds a small Marian model, simulates the conflicting inputs and verifies that training progresses without raising the exception.

Motivation
This pattern appears in real-world usage when composing an encoder from one model and a Marian (or similar) decoder while applying PEFT/LoRA. Making the Trainer resilient avoids forcing each user to patch their own subclass.

Tests
pytest tests/trainer_seq2seq_test.py -q passes locally.

Additional notes
Happy to adjust the location of the guard (e.g. move it to the base Trainer) if reviewers prefer.

Who can review?
Tagging trainer maintainers for visibility: @zach-huggingface @SunMarc

[Rocketknight1]

It's unclear that this fixes the issue! Please don't run code agents on user issues when the cause hasn't been fully identified, they have a strong tendency to fix the wrong thing.