Fix synchronization bug for GPU stream async CPU work

[awni]

Fixes synchronization bugs for asynchronous CPU work that happens in the GPU stream.

The basic premise is that the output events were getting signaled before all the work in the GPU stream actually completed.

For example:

a = mx.ones(5,) # a event value 1
x = mx.load() # x has event value 2
z = mx.all_sum(a) # wait on event value 1

The load would signal the event value to 2 without ensuring that a was actually finished causing the all_sum to start prematurely.

The fix is to simply make a new event for primitives like Load and the distributed ops that need to synchronize with the GPU.

[angeloskath]

Perfect! That would have gone unnoticed for long...

[angeloskath]

That is pretty tricky. Many comm backends have a synchronous/blocking and non-blocking version but I feel that this may be a bit much for now. So I think going without an event is better as it will allow us to better pipeline multiple sends or computation after send in general.