Fix/1616 pd na in by label count

[ugbotueferhire]

Fixes #1616.

Description

DataSummaryPreset and DataDriftPreset crashed with a pydantic.ValidationError whenever a categorical column contained pd.NA — pandas' native missing-value sentinel for nullable dtypes (string, Int64, boolean, etc.). Users had to convert pd.NA to None before passing data in, defeating the point of running exploratory metrics on raw data.

This PR makes both presets handle pd.NA transparently. Missing values are treated as missing — counted by MissingValueCount, never surfaced as a phantom <NA> label in ByLabelCountValue. The fix is three small, layered changes (one direct cause + two defensive guards so the same class of bug can't quietly resurface in future ByLabelCount metrics). No public API changes; no Label type change.

import pandas as pd
from evidently import DataDefinition, Dataset, Report
from evidently.presets import DataSummaryPreset

df = pd.DataFrame({"a": pd.Series(["x", "y", "z", pd.NA], dtype="string")})
ds = Dataset.from_pandas(df, DataDefinition(categorical_columns=["a"]))
Report([DataSummaryPreset()]).run(current_data=ds)
# Before: pydantic.ValidationError on ByLabelCountValue (counts/shares __key__)
# After:  runs cleanly; pd.NA is treated as missing, not as a label.

The same failure hit DataDriftPreset. Both presets now pass.

Root cause

Three cooperating defects:

1. UniqueValueCountCalculation._all_unique_values used pd.Series.unique(), which preserves pd.NA / np.nan. The NA sentinel was then carried into the result dict as a label key.
2. ByLabelCountValue lacked the @validator(..., pre=True) key-coercion that its sibling ByLabelValue has, so non-Label-typed dict keys (e.g. np.int64, pd.NA) reached pydantic uncoerced.
3. convert_types relied on np.isnan(val), which raises TypeError: boolean value of NA is ambiguous on pd.NA.

Changes

• src/evidently/metrics/column_statistics.py — UniqueValueCountCalculation._all_unique_values now calls .dropna().unique(), aligning the unique-value set with the existing value_counts(dropna=True) on the count side. Missing values continue to be reported separately by MissingValueCount (no double-counting).
• src/evidently/core/metric_types.py
  • Add @validator("counts", "shares", pre=True) on ByLabelCountValue, mirroring ByLabelValue.convert_labels.
  • Identity-check pd.NA in convert_types and normalize to None, consistent with Label = Union[StrictBool, int, str, None]. np.nan behavior is preserved (still falls through to pydantic's str-coercion path, matching the existing test contract).

Tests

• New tests/future/presets/test_dataset_stats_na.py — end-to-end coverage matching the issue's exact reproduction on both DataSummaryPreset and DataDriftPreset.
• Extended tests/future/metrics/test_unique_value_count.py — cases for pd.NA in string / object / Int64 dtypes.
• Extended tests/future/test_metric_types.py — parametrize cases covering pd.NA and np.int64 keys on ByLabelCountValue.