Fix PandasArrayExtensionDtype._metadata: must be a tuple, not a string

[saket3395]

Summary

Fixes #8375 — PandasArrayExtensionDtype._metadata is a plain string instead of a tuple, causing AttributeError: 'PandasArrayExtensionDtype' object has no attribute 'v' whenever pandas compares two instances of this dtype (e.g. during a boolean-mask filter on a DataFrame with an array-feature column read back from parquet).

Root cause: pandas.api.extensions.ExtensionDtype expects _metadata to be a tuple of attribute names used for equality comparison. As a plain string "value_type", pandas iterates over it character-by-character ('v', 'a', 'l', 'u', 'e', ...) when comparing dtypes, and crashes looking up an attribute named 'v'.

_metadata = "value_type"

→

_metadata = ("value_type",)

Verification

Reproduced the underlying bug in isolation — a minimal ExtensionDtype subclass using only pandas, no datasets import required:

import pandas as pd
from pandas.api.extensions import ExtensionDtype

class BadDtype(ExtensionDtype):
    _metadata = "value_type"
    name = "bad"; type = object; na_value = None
    def __init__(self, value_type): self.value_type = value_type
    @classmethod
    def construct_array_type(cls): return None

class GoodDtype(ExtensionDtype):
    _metadata = ("value_type",)
    name = "good"; type = object; na_value = None
    def __init__(self, value_type): self.value_type = value_type
    @classmethod
    def construct_array_type(cls): return None

BadDtype("float64") == BadDtype("float64")   # AttributeError: 'BadDtype' object has no attribute 'v'
GoodDtype("float64") == GoodDtype("float64") # True

This confirms the string-vs-tuple distinction is the actual cause of the reported AttributeError, and that the tuple form resolves it, independent of anything else in datasets.

Test plan

• Reproduced the exact failure mode (AttributeError: no attribute 'v') in isolation and confirmed the fix resolves it.
• Didn't run the full datasets reproduction script from the issue (parquet round-trip + Array2D features) in this environment — didn't want to pull in the full datasets/pyarrow stack just to verify a one-character change whose root cause is fully isolated above. Happy to iterate if CI surfaces anything additional.

[kohankhaki]

This is a copy of a PR I have already created: #8376