`Buffer` deallocation always happens on default stream

[shwina]

There are two ways a buffer can be deallocated.

• Explicitly via calling Buffer.close()

  cuda-python/cuda_core/cuda/core/experimental/_memory.py

  Line 63 in ae1cbcf

  def close(self, stream: Stream = None):

• Implicitly via finalizer

  cuda-python/cuda_core/cuda/core/experimental/_memory.py

  Line 43 in ae1cbcf

  weakref.finalize(buffer_obj, self.close)

In the implicit case, no stream= argument is passed to the finalizer function, which means that if a buffer gets refcounted out or gc'd, it will always be deallocated on the default stream.

Possibly we want to change this so that the stream passed to the mr.allocate() function is used for deallocation by default.

[leofang]

In a number of internal discussions this issue has been brought up over and over (most recently on Sep 8 when discussing #674). We decided to align with cccl-runtime and change this to deallocate on the allocation stream, with the understanding that

1. no choice is perfect, and this choice is to serve the most common use cases (we think!) where a short-lived buffer is allocated/deallocated on the same stream
2. users are still encouraged to call .close(stream) (cuda.core) or .destroy(stream) (cccl-runtime) to preserve stream ordering explicitly
3. this is safer when we reimplement MRs on top of those from cccl-rt