More donation take 2

[awni]

This is a redux of #1858

• The race condition of checking use counts is mitigated since the checks are done on the main thread now
• Use weak pointers to hold references to other inputs to avoid breaking async eval. Another option is to just use a more restrictive policy when doing async eval and avoid this complexity.
• Also moves all the buffer retention logic from the CPU / GPU task submission to the main thread since it's all duplicated now.

Comment from #1858 still applies (copied below):

Previously if an input didn't get donated it was held by the command buffer and no longer eligible for donation.

So sequences like:

a = x + y
b = mx.abs(x)

Assuming x didn't get donated to a then it is impossible for it to be donated to b.

This changes that by removing buffers that are safe to not hold in the command buffer.