perf: build IdxInsert keys from live registers instead of re-seeking (#663)

[iheitlager]

Summary

• INSERT and UPDATE codegen already hold every new/updated column value in col_regs/rowid_reg before writing the row; the secondary-index-key builder (emit_index_key_ops) instead SeekRowid'd back onto the just-written row and re-read those same values via Opcode::Column/Opcode::Rowid — an unnecessary btree seek + page touch per row on every INSERT/UPDATE against an indexed table.
• Adds emit_index_key_ops_from_regs in src/codegen/index_maintenance.rs, which builds the same key layout via Opcode::Copy straight from the existing value registers, and switches the IdxInsert call sites in insert.rs/update.rs to it, dropping their SeekRowid.
• IdxDelete call sites (removing a different, already-on-disk row's stale entries — UPDATE's old-row cleanup, and the INSERT OR REPLACE conflict paths) are untouched — they have no pre-existing register run for that row and must stay cursor-based.

Refs: emit_index_key_ops doc comment's stated rationale ("no register-copy opcode in the frozen V2 set") was factually wrong — Opcode::Copy exists and is used in 15+ other codegen sites.

Test plan

• cargo test --test corpus index_maintenance — all 12 scenarios pass (oracle PRAGMA integrity_check + indexed-lookup verification)
• cargo test --test corpus --test vdbe_write_opcodes --test vdbe_cursor_positioning --test vdbe_cursor_sorter — 380 + opcode-shape tests pass
• cargo test --test codegen_insert --test codegen_update --test parity — pass
• cargo clippy --all-targets — clean
• Manual opcode dump confirms the fix: post-Insert, the program now emits two Copys straight into IdxInsert with no SeekRowid/Column re-read
• cargo bench --bench crud -- insert_batch_10 ran (high variance from a concurrent bench in another worktree during this run, but no regression)

spend: matched estimate (medium)

Closes #663