fix: bound IVF-Flat serialization allocations

[jerry-024]

Summary

Bound IVF-Flat serialization memory usage by streaming vector payloads through one reusable write buffer instead of materializing full sorted-vector and encoded-byte buffers for each list.

The serialized output remains byte-for-byte compatible with the existing IVFL v1 reader and is independent of write chunk boundaries.

Changes

• Reuse a private 64 MiB write buffer for centroid and inverted-list vector encoding.
• Preserve row-ID sorting while encoding referenced vectors in bounded chunks.
• Split oversized vectors across writes so the reusable buffer and vector payload writes stay within 64 MiB.
• Add regression coverage for byte stability, ID/vector ordering, multi-chunk writes, and the write-size bound.

Testing

• cargo fmt --all -- --check
• cargo +1.94.0 test -p paimon-vindex-core ivfflat_chunked_writer
• cargo +1.94.0 test -p paimon-vindex-core --lib (446 passed, 1 ignored)

Notes

• No public API, dependency, index option, or file-format version changes.
• The index vectors, row-ID sort permutations, and encoded IDs remain resident; this change only removes the two full vector-payload-sized serialization allocations.

[shyjsarah]

+1