fix(encoding): reject mismatched miniblock buffer sizes

[lance-gatefixer]

Summary

• derive the checked total represented by mini-block chunk metadata before building byte-prefix sums
• reject metadata/data buffer size mismatches as a contextual corrupt-file error in every build profile
• cover both under-reported and over-reported data buffer sizes

Root cause

build_chunk_index selected the prefix-sum storage width from data_buf_size while accumulating independent sizes decoded from the metadata words. Their equality was enforced only by a debug assertion, so debug builds panicked and release builds could retain offsets beyond the buffer or choose an undersized prefix representation.

The historical producer path in the reported run was removed by the column-stitching revert, but the reader-side integrity boundary remained unchecked.

Validation

• cargo test -p lance-encoding test_rejects_mismatched_miniblock_data_size -- --nocapture
• cargo test -p lance-encoding --lib (1819 passed, 5 ignored)
• cargo test -p lance --lib index::vector::ivf::tests::test_create_index_nulls::case_5_ivf_hnsw_sq::index_version_1_IndexFileVersion__Legacy -- --exact --nocapture
• cargo fmt --all -- --check
• cargo clippy --all --tests --benches -- -D warnings

Fixes #8947

[lance-gatekeeper]

✅ Gate recommendation: approve.

This closes a real reader-side integrity gap: mini-block metadata must now cover exactly the declared data buffer before offsets are cached, so malformed pages fail with a contextual corrupt-file error instead of panicking in debug builds or scheduling invalid ranges in release builds. The validation is localized to the parsing boundary and preserves compatibility with the stable file formats.