Fix KeyError in 4D attention mask creation for encoder-decoder model

[OliBomby]

What does this PR do?

Fixes KeyError in 4D attention mask creation for encoder-decoder models when generating with attention mask and static cache.

This occurs because inputs_embeds does not get added to model_inputs for encoder-decoder models, like Whisper. Subsequently it tries to check if model_inputs["inputs_embeds"] is none and throws a KeyError.

cc @gante

[gante]

@OliBomby good catch! Thank you for opening the PR :D

This and other encoder-decoder issues are being addressed in #38235 and #38108, so I'm not going to merge this PR to avoid creating merge conflicts 🤗 (cc @vasqu)

[OliBomby]

Ah in that case I'll close this PR. Hope to see the others merged soon!