Fix AttributeError in run_classification.py when detecting multi-label data

[shtse8]

What does this PR do?

Fixes the AttributeError: 'List' object has no attribute 'dtype' crash in run_classification.py when loading JSON data with list-type labels for multi-label classification (reported in #43116).

Problem

When label data is in list format (e.g. "labels": ["4359", "5181"]), the HuggingFace datasets library infers the feature type as Sequence/List, which does not have a dtype attribute. The existing code crashes in two places:

1. is_regression check (line 416): raw_datasets["train"].features["label"].dtype in ["float32", "float64"] → AttributeError
2. Multi-label detection (line 442): ...features["label"].dtype == "list" → AttributeError

Fix

1. Use getattr(label_feature, "dtype", None) for the regression check, so list-type features safely return None (not regression)
2. Use isinstance(feature, datasets.Sequence) for multi-label detection instead of checking .dtype == "list"

Reproduction

# With JSON data like: [{"text": "...", "labels": ["a", "b"]}, ...]
# Before fix:
#   AttributeError: 'List' object has no attribute 'dtype'
# After fix:
#   Correctly detects multi-label classification

Note: This PR only fixes the crash in multi-label detection. The separate thresholding issue for empty predictions is tracked in #43135.