Currently Core supports an in-memory mode, but it still temporarily uses a backing memory mapped file (it just never explicitly syncs it to disk). It does this so that multiple processes can concurrently work on the same in-memory dataset.

This is generally a good thing, but for running under wasm, memory mapped files become very problematic, and there is no need for multi-process support there anyways. So it would be very useful to support a "pure" in-memory mode that uses regular allocations rather than memory mappings.