fix: filter stale physical row ids in take

[lance-gatefixer]

Summary

• validate non-stable physical row IDs against the dataset deletion mask before TakeExec reads data
• drop index entries for deleted rows and removed fragments from both the address array and carried input columns
• add regression coverage for stale physical row IDs that previously caused a RecordBatch length mismatch

Root cause

TakeExec::get_row_addrs only built a validity mask when stable row IDs were enabled. With default physical row IDs, stale index entries were passed through unchanged. Fragment reads omit deleted rows, so the taken columns could contain fewer rows than the upstream FTS batch and merge_with_schema failed with unequal RecordBatch lengths.

Validation

• cargo test -p lance io::exec::take::tests (13 passed)
• cargo fmt --all -- --check
• cargo clippy --all --tests --benches -- -D warnings

Fixes #7508

[lance-gatekeeper]

✅ Gate recommendation: approve.

This closes the legacy TakeExec parity gap: physical row IDs are checked against the current deletion mask before carried columns and fetched data are merged, matching the existing stable-ID and modern live-view behavior. Deleted and removed-fragment hits now drop while surviving IDs, scores, and fetched values remain aligned.