fix(lfm2-vl): unwrap InputEmbeddingsFeatures at the language-model boundary

Second defect in the same upstream module: lfm2_vl Model.__call__ forwards
the whole InputEmbeddingsFeatures dataclass as inputs_embeds, while
LanguageModel.__call__ is typed Optional[mx.array] and indexes it —
"InputEmbeddingsFeatures object has no attribute shape". Unwrapping at the
language-model boundary fixes it regardless of which caller built the
features object, and plain arrays pass through untouched.