perf(table): keep clustered row ID ranges compact in index

[Abhisheklearn12]

Summary

PR #9311 introduced compact Ranges row ID segments, but merged RowIdIndex construction expands them into per-row pairs even when there are no deletions. Keep the compact segment and map its positions to a contiguous address range.

Tests cover gaps, shifted physical offsets, overlapping fragment spans, and batched lookups. The deletion-aware path and on-disk format are unchanged.

Fixes #9487.

Benchmark

BuildIndexClusteredRanges uses 100 fragments, 500,000 live IDs across 1,000,000 ID slots, and empty deletion vectors. Fixture construction is outside the timed loop. Both variants used the same benchmark and release-with-debug profile with 30 Criterion samples. The baseline differs only by removal of the new Ranges match arm.

Index construction path	Estimate	Criterion interval
Existing per-row expansion	4.335 ms	4.327 to 4.344 ms
Compact Ranges path	26.44 µs	26.15 to 26.74 µs

This is about 164× faster for this synthetic index-build case. Dataset opening, storage I/O, and query latency were not measured.

Validation

• cargo fmt --all
• cargo clippy --all --tests --benches -- -D warnings
• cargo test -p lance-table rowids:: (131 passed)

[lance-gatekeeper]

✅ Gate recommendation: approve.

This removes the per-row expansion identified in #9487 for compact ranges without deletions. The position-to-address mapping preserves lookups across gaps and shifted offsets, while overlapping chunks retain the existing merge path. The broader BTree replacement has separate prerequisites, so this scoped improvement can land independently.