feat: public APIs for trivial column access (raw byte ranges)

[wkalt]

Add APIs that let downstream consumers read raw value bytes directly from the file without going through the decoder pipeline, when the column encoding is simple enough (non-nullable, fixed-width, flat, uncompressed).

New APIs:

• ColumnInfo::is_trivially_decodable() — true when on-disk bytes match the in-memory representation. Handles MiniBlock, FullZip, and legacy v2.0 flat encodings.
• ColumnInfo::trivial_page_descriptors() — returns per-page metadata (buffer offsets, bits_per_value, has_large_chunk) for locating raw data in the file.
• miniblock_raw_data_ranges() — parses miniblock chunk metadata words to produce (offset, length, num_rows) tuples pointing past headers to raw value data.
• PageInfo::miniblock_layout() / full_zip_layout() — extract the specific layout variant from the encoding protobuf.
• CachedFileMetadata::column_index(name) — maps a column name to its index in column_infos, handling FSL child resolution.

These enable a speedup for brute-force KNN when vector data is cached in memory, by allowing the caller to reinterpret cached pages as &[f32] without Arrow allocation or DataFusion plan execution.

[github-actions]

PR Review: feat: public APIs for trivial column access (raw byte ranges)

Missing Tests (P0)

Per project policy: "All bugfixes and features must have corresponding tests. We do not merge code without tests."

This PR adds ~250 lines of new public API surface with zero tests. At minimum, the following should be tested:

• is_trivially_decodable() — true for simple flat/miniblock/fullzip pages, false for nullable, dictionary, compressed, sub-byte, multi-buffer cases
• trivial_page_descriptors() — correct offsets, bits_per_value, page_row_offset accumulation, has_large_chunk propagation
• miniblock_raw_data_ranges() — correct parsing of both u16 and u32 meta words, correct data_offset/data_len/num_values for multi-chunk pages, and correct last-chunk remainder handling
• column_index() — scalar column, FSL column, missing column

Silent continue in trivial_page_descriptors (P1)

In trivial_page_descriptors(), the method first checks is_trivially_decodable() (which validates all pages), but then has continue branches in the per-page loop for pages that don't match any known layout. If is_trivially_decodable and the per-page logic ever drift apart, continue will silently produce fewer descriptors than there are pages, and the page_row_offset accumulation will be wrong. Consider making this an error/None return rather than silently skipping, or removing the redundant checks since is_trivially_decodable already guarantees the structure.

CHUNK_HEADER_BYTES should be derived, not hardcoded (P1)

CHUNK_HEADER_BYTES: u64 = 8 happens to be correct today for trivial miniblocks (2B num_levels + 2B or 4B buffer_size, padded to 8-byte alignment), but this is a fragile implicit coupling with the miniblock encoding format. If the format ever changes (e.g., different alignment, additional header fields), this constant will silently produce wrong offsets. At minimum, add a comment explaining the derivation. Better yet, reuse the existing header-parsing logic or assert the invariant.

column_index relies on .last() (minor)

column_index returns proj.column_indices.last() with the comment "handles FSL child resolution." This works because from_column_names returns parent + child indices for composite types. A brief doc comment explaining why .last() is the right choice (the last index is the leaf data column) would help future readers.

[wkalt]

this is proof of concept for a fast path to KNN scanning in-memory lance data pages. Not expected to merge or be the approach we ultimately take.