Instead of updating shared texture arrays directly:

• Upload texture data to long lived buffers that aren't necessarily grouped by sampler types.
• Have a separate set of textures into which images are packed with a very aggressive eviction policy for batching purposes.
• Sll of these textures don't need to be texture arrays.

The advantages are:

• Flexibility: we could use different packing strategies for the staging and batching textures.
• We don't need to worry about resizing the staging textures and the aggressive eviction of the batching textures means they can also remain small and not change size.
• Allow using more efficient packing algorithms without worrying too much about fragmentation of the batching texture atlases.
• Get rid of texture arrays in the texture cache.

The main disadvantage is that many on-screen images exist twice causing an increase in GPU memory and some extra copying work between the staging and batching textures.

It's a bit unclear whether there is more we can/should do about packing and updating staging textures beyond what we do with the shared texture atlases today.