Parquet: Variant shredding follow-ups from PR #14297

[nssalian]

Changes

Follow-ups from the PR #14297 summary thread and some additional cleanup

1. Added javadoc to TIE_BREAK_PRIORITY so readers know what the constant is for. Updated the class-level javadoc to remove the stale TreeMap reference and link to TIE_BREAK_PRIORITY.
2. Moved PathNode.objectChildren from TreeMap to HashMap. Alphabetical schema field order is preserved by sorting once in createObjectTypedValue at schema build time.
3. Added debug logging in ParquetFormatModel.buildShreddedAppender that records the buffer size at construction and, per inference flush, the buffered row count and the inferred shredded field count.
4. Reused the existing isDecimalType helper in observe and dropped the now-unused VariantPrimitive import.
5. Cached the result of getMostCommonType per FieldInfo. The schema build calls it once at the root and again per node, currently rebuilding the same family-aggregation every time.
6. Used a min-heap of size MAX_SHREDDED_FIELDS for the field cap, replacing the full sort plus n-sized intermediate ArrayList plus HashSet allocation.
7. Used an int[] for FieldInfo.typeCounts keyed on PhysicalType.ordinal(), replacing Map<PhysicalType, Integer>. Removed the lambda capture allocated by Map.compute on every observation and the boxing of Integer.

Test plan

• testIntermediateFieldCapLimitsTrackedFields extended with three assertions verifying the min-heap retains the alphabetically-earliest fields when counts are tied.
• New testUuidFieldIsTrackedAndShredded exercises int[] indexing for a high-ordinal PhysicalType value.
• Build passed locally and TestVariantShreddingAnalyzer as well.

[huaxingao]

LGTM

[huaxingao]

Thanks @nssalian for the PR!