Fix Parquet 0.8.0 type fidelity and selective reads

[perNyfelt]

Implement matrix-parquet/req/v0.8.0-fixes.md in matrix-parquet, keeping 0.8.0-SNAPSHOT and the existing unreleased release section.

• Correct unsigned integers, INT96 and legacy MAP reads; preserve nested Matrix leaf types through allowlisted matrix.nestedTypes metadata.
• Use exact residual filters for unrepresentable predicate bounds while retaining safe partial AND pruning. Default all timestamp classes to MICROS truncation, with overflow-safe NANOS opt-in.
• Fix nested shape inference and error paths, bloom/codec validation, secure stream spooling, dataset names, glob depth/separators and weak-ETag preconditions. Remove dead reader/POJO paths and update documentation.

The 200-column × 50,000-row write measurement improved from 9,590 ms to 3,427 ms with identical output size (one local run per implementation).

Verification passed:

• ./gradlew :matrix-parquet:codenarcMain :matrix-parquet:codenarcTest --no-daemon --rerun-tasks --console=plain
• ./gradlew :matrix-parquet:spotlessCheck --console=plain
• ./gradlew :matrix-parquet:test --console=plain
• PARQUET_PYTHON=/tmp/matrix-parquet-fixtures/bin/python ./gradlew :matrix-parquet:test :matrix-parquet:compileConsumerApi -PrunSlowTests=true --console=plain — final module run: 156 passed, 6 optional skips.
• PyArrow 21.0.0 verify_matrix_output.py: SNAPPY/ZSTD and CDC rewrite/readback.
• ./gradlew :matrix-parquet:test --tests '*ParquetReadBenchmarkTest' -Pparquet.baseline=true -Pparquet.benchmarkRows=10000 --rerun --console=plain — 1 executed, 0 skipped.
• ./gradlew :matrix-parquet:test --tests '*ParquetLayoutBenchmarkTest' -Pparquet.layoutBenchmark=true --rerun --console=plain — 2 executed, 0 skipped.
• ./gradlew test -Pheadless=true --console=plain — full repository suite passed, including a final rerun.

The completed requirement file contains the detailed verification record, including benchmark setup and a clean rerun after an internal CodeNarc/Groovy worker error. No tags were created.

Windows parsing rules were unit-tested on macOS; no Windows host run was performed. Weak ETags and legacy MAP layouts were verified with local HTTP/synthetic Parquet fixtures, without production-server or real legacy-file validation.