Problem: memiavl has relatively big gc pressure

Depending on the traffic of the chain and interval of snapshot rewriting, memiavl need to manage lots of small memories for `MemNode`s and key/value slices, which could put a big pressure to golang GC, it might make sense to use some manual memory management.

## Memory Arenas
One obvious solution is [memory arenas](https://uptrace.dev/blog/golang-memory-arena.html) feature introduced in go 1.20, we just allocate these objects with big chunks of arenas.

### Negatives

- we can't reclaim and reuse memory, arenas don't support deallocating, before dropped as a whole when we switch snapshot, it's maybe alright assuming deletion happens more rarely than insertion/updates.

### Manual Memory Management

- Implement `MemNode` as a `[]byte`, similar to the nodes in mmap-ed snapshots, allocate in a big buffer, maintain an intrusive free list to reuse nodes.
- Leave key/value slices to golang GC.

Start by profiling memiavl under representative chain traffic and snapshot rewriting, focusing on MemNode and key/value slice allocations. Compare the proposed memory-arena approach with manual MemNode allocation and reuse, while leaving key/value slices to Go's GC. Done should include an evidence-backed implementation direction and measured GC impact.

Written by the indexing model from the issue text.