Introduce sparse buffer uploads for mesh input uniforms.

[pcwalton]

Currently, all mesh input uniforms (primarily the transform data) for all mesh instances are uploaded to the GPU every frame, even for those meshes that haven't changed their transform. This has become a bottleneck when scaling to millions of mesh instances.

This commit solves this issue by introducing sparse buffers. A sparse buffer is a buffer that tracks which elements are changed and, if the number of changes is small enough, pushes only the elements that have changed to the GPU and uses a compute shader to scatter those elements into their proper positions in the buffer. This supports the access patterns of large worlds, in which there may be millions of static mesh instances but, on any given frame, only a few thousand change their transform.

Although this PR only makes the mesh input uniform buffer sparsely updated, I want to use this infrastructure to store bins in the future, in order to optimize the other operation that must currently send per-mesh-instance data to the GPU every frame. That data doesn't need to be written from multiple threads, and so at that time we'll want to introduce a SparseBufferVec to go alongside the
AtomicSparseBufferVec that this patch introduces. In preparation for that, I've factored most of the functionality that will be common to SparseBufferVec and AtomicSparseBufferVec into separate functions in the sparse_buffer_vec module. That will allow the future bin slab work to introduce SparseBufferVec without having to refactor the work in this patch.

On many_cubes --instance-count 1000000 --no-cpu-culling, this reduces the median time spent in write_batched_instance_buffers from 12.4 ms to 0.876 ms, a 14x speedup.