Optimize compute dispatch and buffer upload

[tychedelia]

Currently, every compute dispatch runs a full app.update() to be maximally pessimistic about changes to GPU resources that may not have been flushed to the GPU yet. For many particle sims which chain dispatches together this is exceedingly wasteful. This adds a new compute_dispatch_no_update which simply does the dispatch if the user is sure that there have been no un-flushed writes in the meantime.

Also changes how we do buffer uploads. Previously, we would block on a full readback to sync the buffer into CPU memory. Now, we just write directly into the GPU buffer. This is fine, because reads still incur the full sync but if the user only writes but never reads we can stay on the fast path and let the CPU side buffer get out of sync.

[catilac]

awesome! what is a good way to test this?

[tychedelia]

awesome! what is a good way to test this?

great q, i think this is an example of where the tests passing are a good indication. this should mostly be a behind the scenes thing (in a few cases) or a new performance tool for power users. i'm going to open a pr that reworks some of the flocking stuff soon which may be a good place to dig deeper too.

[tychedelia]

Reopened from the upstream branch so it can anchor a PR stack.