fix: handle ragged batch inputs in Qwen2_5_VLProcessor mm_token_type_ids computation

[s-zx]

What does this PR do?

Fixes a crash in Qwen2_5_VLProcessor.__call__ when processing batched inputs without padding (padding=False).

Root cause: When the tokenizer returns sequences of different lengths (ragged list), np.array(text_inputs["input_ids"]) creates an object array instead of a 2D integer array. The subsequent element-wise comparisons (== image_token_id, == video_token_id) then fail or produce wrong results.

Fix: Detect batch inputs (list-of-lists) and process each sequence individually, which correctly handles both padded (uniform length) and unpadded (ragged) batches.

Before:

array_ids = np.array(text_inputs["input_ids"])  # fails for ragged batches
mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
mm_token_type_ids[array_ids == self.image_token_id] = 1

After:

if isinstance(input_ids, list) and input_ids and isinstance(input_ids[0], list):
    # process each sequence individually
    for ids in input_ids:
        arr = np.array(ids)
        ...

Fixes #44514

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: qwen2_5_vl

[zucchini-nlp]

Will close in favor of #44563 (comment) which solves the issue for all models at once