feat: Collect Parquet NaN metrics during writes

[WZhuo]

Collects NaN value counts for float and double columns during Parquet writes, since the Parquet footer statistics do not track NaN counts.

Changes

• Write-side NaN metric collection (FieldMetricsCollector): A visitor that walks each record batch before writing, accumulating value counts, null counts, NaN counts, and NaN-excluding lower/upper bounds for float/double fields.
• MetricsConfig-aware skipping: Fields whose MetricsMode is kNone are skipped entirely, avoiding wasted work.
• Integration with existing footer metrics: Write-side FieldMetrics take precedence over footer statistics in ParquetMetrics::GetMetrics, so NaN counts are populated while counts/bounds still fall back to footer stats when write-side data isn't available.
• Tests: ParquetMetricsTest now overrides ReportsNanCounts() to true, and existing NaN test cases verify NaN counts alongside existing value/null count assertions.

Behavior alignment with Java

• Fields nested inside lists/maps do not get NaN metrics (both Java and C++ agree — Java collects then discards; C++ skips collection entirely).
• NaN values are excluded from lower/upper bounds in both implementations.
• Float/double fields with all-NaN values correctly set nan_value_count without setting bounds.

[wgtmac]

Non-blocking: this could be made cheaper by carrying Arrow validity bitmaps instead of materializing a std::vector<uint8_t> and calling array.IsNull(i) for every row at each struct level. The main detail is that this needs to preserve the current parent-validity behavior: child arrays of a null StructArray do not necessarily have those parent nulls reflected in their own null bitmap, so the effective validity should be parent_validity AND array.null_bitmap().

One possible approach is to keep a bitmap/buffer plus offset for the current effective validity, and when descending into a field, use Arrow bitmap utilities to AND it with the child array validity bitmap (or just reuse the parent bitmap when the child has null_count() == 0). Then the float/double collector can iterate only the effective bitmap bits instead of indexing a byte vector. This should avoid per-level byte-vector allocation and reduce the per-row validity checks for wide/deep struct writes.