memiavlstore.Store: queries can race ApplyChangeSet mutating the live *memiavl.Tree

### What

`memiavlstore.Store.tree` is an `atomic.Pointer[memiavl.Tree]`, swapped by `SetTree` after a commit builds the next version. That makes the pointer swap itself race-free, but it doesn't protect the tree object underneath.

### Issue

`rootmulti.Store.flush()` calls `rs.db.ApplyChangeSets(...)` (store/rootmulti/store.go), which mutates the current `*memiavl.Tree`'s internals (node cache, dirty tracking) in place, before `SetTree` publishes the new tree. A query goroutine on another ABCI connection can `Load()` the pre-commit tree pointer and call `Get`/`Has`/iterate on it while `ApplyChangeSet` is mutating the same object — a data race on the tree's internals, not on the `Store.tree` field.

### Repro sketch

- Goroutine A: `rootmulti.Store.Commit()` -> `flush()` -> `rs.db.ApplyChangeSets(...)` mutating tree T in place.
- Goroutine B (concurrent ABCI query on another connection): `memiavlstore.Store.Get(key)` -> `st.tree.Load()` returns T (not yet swapped) -> reads T's internals concurrently with A's mutation.

`go test -race` across a workload that mixes queries with commits should be able to surface this.

### Notes

Found while reviewing #103 (WAL durability fix) — noted in a comment at `store/memiavlstore/store.go` on the `tree` field, but out of scope for that PR. Filing separately to fix later.

Start with store/memiavlstore/store.go, especially the tree field and its Get/Has methods, then trace rootmulti.Store.flush() in store/rootmulti/store.go through ApplyChangeSets and SetTree. Run a mixed query/commit workload with go test -race; done means concurrent queries and commits no longer race on the live tree internals.

Written by the indexing model from the issue text.