Support BSI FileIndex and indexed primary-key and data-evolution writes

[JingsongLi]

Summary

• Add Java V1-compatible BSI FileIndex reading and writing for integral, date, time, decimal, and timestamp values. Verify Rust output byte-for-byte against a Java-generated fixture, including signed values and nulls.
• Generate FileIndex metadata for primary-key KV files, postpone-bucket files, and data-evolution base and partial-column files. Index positions follow physical file order after sorting and merging; partial-column indexes remap field positions to their file schema. Changelog files remain unindexed.
• Use primary-key-only indexes before sort-merge and indexes on independent data-evolution files before decoding. Preserve residual filtering, row ranges, and deletion-vector behavior. Use the normal commit path so indexed dynamic-bucket writes retain HASH index and input-changelog metadata.
• Cover embedded and sidecar indexes, rolling files, abort cleanup, multiple snapshots, sorted chunk boundaries, dynamic and postpone buckets, malformed BSI payloads, and Java interoperability.

Verification

• cargo test -p paimon --lib (3136 passed, 6 ignored)
• cargo clippy -p paimon --lib -- -D warnings
• cargo fmt --all -- --check
• git diff --check

[leaves12138]

LGTM. Reviewed a88663f; no blocking issues found.

Checked the BSI V1 encoding/value mapping against the Java implementation and the included Java fixture, signed/null and temporal handling, physical index row positions after sorting/merging, partial-file field remapping, and preservation of HASH/input-changelog metadata on the normal commit path. Read-side pruning remains limited to primary-key predicates before sort-merge and independent data-evolution files, preserving residual filtering and column-merge semantics.

Local validation (default features):

• cargo test --locked -p paimon --lib: 3136 passed, 6 ignored, including the 10 BSI and 28 file-index writer tests.
• incremental_batch_scan_test and incremental_diff_extended_test: 39 + 6 passed.
• Two additional temporary regression tests passed: BSI selections combined with snapshot-specific DVs, small read batches, projection and Diff/audit reads in both DV modes; data-evolution nonzero row-id offsets, intersected row ranges and updated partial-column values. Both exercised embedded and sidecar indexes with index reads enabled and disabled.

CI has no reported failures at the time of review; three jobs are still running. Optional feature builds were not rerun locally.