What does this PR do?

Fixes #46328.

SequenceFeatureExtractor.pad() normalizes its inputs with:

for key, value in processed_features.items():
    if isinstance(value[0], (int, float)):
        processed_features[key] = to_numpy(value)
    else:
        processed_features[key] = [to_numpy(v) for v in value]

When value is already a batched numpy array, the else branch rebuilds it into a
Python list of per-example arrays ([to_numpy(v) for v in value]). That iteration and
per-row copy is pure overhead and becomes very slow for large inputs (the issue reports
several minutes on a 25-minute audio file).

It is also unnecessary: the downstream _truncate/_pad logic only ever indexes
value[i], which behaves identically for a list of arrays and for a batched ndarray
(same per-example shapes, same len() for the batch dimension). So an already-batched
array can be used as-is.

This PR skips the conversion when value is already an np.ndarray. The common
list-of-arrays / list-of-tensors path is unchanged, so existing behavior is preserved.

Before submitting

• This PR fixes a reported issue (SequenceFeatureExtractor.pad wasting time converting numpy array to list of numpy arrays #46328).
• Did you read the contributor guideline?

Generated with Claude Code (AI-assisted, human reviewed).