fix: IndexCursor::seek does O(log n) tree descent instead of linear scan

[iheitlager]

Summary

• IndexCursor::seek() (src/btree/index.rs) previously called first() then walked forward via next(), fully decode_record-ing every candidate until finding one >= target: O(n) cells scanned and eagerly payload/overflow-reassembled per seek. TableCursor::seek already binary-searches cell pointers per tree level; IndexCursor never got the equivalent treatment.
• seek() now descends the tree: at each level it binary-searches that page's cell array for the leftmost cell >= target, descends into that cell's left child (recording the cell as a fallback via the frame's step, exactly mirroring how ordinary first()/next() traversal would leave it), and falls back to rightmost when no cell qualifies. A final advance() call resolves the right entry using this same stack state — the leaf match if there is one, or the nearest ancestor's fallback cell, or None. This means only the O(log n) cells actually probed per level get decoded, and continued next() calls behave identically to full-scan traversal.

Test plan

• Added secondary_index_seek_between_keys_matches_full_scan_and_next_continues — seeks a value between two stored keys (exercising the interior-cell fallback path), cross-checks against a full scan, and confirms next() continues correctly from the seek position.
• Added secondary_index_seek_past_every_key_returns_none and secondary_index_seek_before_every_key_returns_first_row for the boundary cases.
• Existing secondary_index_seek_matches_oracle, secondary_index_walk_matches_oracle_binary_order, secondary_index_last_prev_matches_reversed_forward_walk all still pass unchanged.
• cargo test --release — full suite green (979+ tests, 0 failed).
• cargo clippy --all-targets — clean.

Closes #661

spend: ~1x estimate