Generalizing the ChunkManifest class to hold native chunks as well as virtual refs would unlock several features.

It's needed for:

1. Writing an IcechunkParser (Reading virtual references back out into VirtualiZarr Manifests earth-mover/icechunk#104)
2. Concatenating virtual and non-virtual data in memory (Icechunk already supports doing this on-disk via append_dim)
3. ManifestStore.to_icechunk()/kerchunk(), and thereby making xarray an optional dependency (Make xarray an optional dependency? #521)

The problem is that the current implementation of ChunkManifest uses a clever trick: it's just 3 numpy arrays in a trenchcoat. This gives us loads of stuff for free:

• Efficient contiguous in-memory representation
• Efficient handling of Variable-length strings (via the numpy 2 dtype)
• Efficient functions for iterating over every element
• Efficient multi-dimensional concat/stack functions for merging chunk manifests
• No unusual or non-python dependencies

But I don't know how to keep that design and also store non-virtual chunks in arbitrary locations within those arrays.

Some alternatives:

• Numpy object array (almost certainly very inefficient)-
• Keep the 3 numpy arrays, but have an auxiliary buffer in which to store non-virtual chunks in memory (proposed on Support in-lined Kerchunk references #794 by @maxrjones)
• Write our own in-memory chunk manifest class in rust