perf: optimize filtered vector search pre-scan and low-selectivity path

[anoop-narang]

Summary

• Narrow pre-scan projection: Pre-scan now projects only _key + columns referenced by filter expressions, instead of all scalar columns. Reduces bytes scanned from ~49MB to ~2.6MB for a 1.2M row dataset.
• Index-get low-selectivity path: Replace the full Parquet scan (which read all columns including vectors) with index.get(key) to retrieve vectors directly from USearch by key. Computes distances for only the valid keys, then fetches result rows from the lookup provider. Eliminates the 7GB vector column scan entirely.
• Dead code cleanup: Remove parquet_native_execute, ScoredRow, compute_distance_for_row, and full_scan from SearchParams — all superseded by the index-get path.

Performance impact (1.2M row dataset, 768-dim vectors)

Query type	Before	After
WHERE filename = '...' (pre-scan)	49MB / 54ms	2.6MB / 8ms
WHERE title LIKE '%..%' (low-sel)	6.95GB / 8s	~0ms (40 index.get calls)

Test plan

• All 40 existing tests pass (optimizer rule, execution, parquet-native paths)
• Verify WHERE filename = '...' filtered vector search returns correct results
• Verify WHERE title LIKE '%..%' filtered vector search returns correct results and is fast
• Verify unfiltered vector search is unaffected

[claude]

Using .ok() here silently converts a missing key column into None, which means _key would be excluded from the projection. The error surfaces later at execution time as a confusing column-not-found or wrong-column panic rather than a clear planning error. Suggest being explicit:

Suggested change

	let key_col_idx = scan_schema.index_of(&registered.key_col).ok();
	let key_col_idx = scan_schema.index_of(&registered.key_col).ok();

Or, for an earlier, clearer failure:

let key_col_idx = scan_schema
    .index_of(&registered.key_col)
    .map_err(|_| DataFusionError::Execution(format!(
        "USearchExec: key column '{}' not found in scan schema",
        registered.key_col
    )))?;

(Leaving the .ok() is fine too — the failure will still surface at execution time. Just a suggestion for debuggability.)

[anoop-narang]

Good catch — adopted the explicit error. Now fails at planning time with a clear message instead of silently excluding _key from the projection.