Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@benches/merge.rs`:
- Around line 9-11: The benchmark creates multiple Memtable instances with the
same ID (Memtable::new(0_u64)), violating the MemtableId uniqueness invariant;
change both maps that build memtables (the one assigning to variable memtables
and the second similar loop around lines 44-46) to supply distinct IDs (e.g.,
Memtable::new(i as u64) or a running counter) so each Memtable gets a unique
MemtableId instead of 0.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@benches/tree.rs`:
- Around line 16-23: The TempDir guard is being moved into Config::new which
drops the guard before .open() runs; update each Config::new call (e.g., the one
shown and the other setup blocks) to pass the Path reference instead of the
TempDir by using path.path() or folder.path() so Config::new receives a path
slice. Search for all Config::new(...) occurrences in benches/tree.rs (lines
around the shown blocks and the other listed blocks) and replace the first
argument from the TempDir variable (path or folder) to path.path() /
folder.path(); leave the rest of the call (SequenceNumberCounter::default(),
.use_cache(...), .open().unwrap()) unchanged.
- Around line 404-420: The benchmark inserts a large value into the tree but
doesn't flush, so the timed "blob tree get" measures memtable lookup rather than
blob storage; before starting the c.bench_function closure call the tree's flush
method (e.g., invoke tree.flush and handle/unpack its Result) to persist the
blob to storage, ensure the flush completed successfully, then run the timed
b.iter to measure blob read performance; reference:
Config::new/...with_kv_separation, tree.insert, and the "blob tree get" bench.