Consider using offset-allocator to pack multiple effects into the same buffer

[pcwalton]

Bevy 0.15 uses https://github.com/pcwalton/offset-allocator to pack vertices and indices from separate meshes into a small number of GPU buffers. This saves on wgpu overhead because it doesn't have to switch bindings so much. We should consider doing this.

[djeedai]

I looked briefly @pcwalton, and although that looks neat for vertices and indices, that's sounds pretty useless for anything else, unless I missed something. The reason is that it doesn't guarantee alignment, so you can't use it to sub-allocate a storage buffer, because if you don't align to min_storage_buffer_offset_alignment then you can't bind the sub-allocation directly. You'd have to use dynamic indexing inside the shader for that (bindless style), which is a lot more complex to handle. But agreed for vertices and indices though.

[djeedai]

Note for self: looking at the bin table, it looks like for any allocation size greater than or equal to align * 8, then all bins are aligned on align. That means for any allocation of at least 256 bytes (align=32, e.g. NVIDIA) or 2048 bytes (align=256, macOS/Metal), we can use the allocator. Otherwise the problem is trivial, so we can manually round up to the nearest multiple before asking the allocator.