Added the following logic:
* in `enqueue_nd_range_kernel`, we catch all generated events for each
queue.
* those events get cleared when calling `queue.finish` or
`cl.wait_for_events`.
* if a kernel argument's queue's events are not in the current `wait_for`,
we complain.

Uses WeakSet to store events so we don't have to worry when they get
destroyed.

The check for out of sync events is also more strict: we now store the
event per (buffer, op) as well as per queue and only check events that
are still valid in both and only for the current buffer.

# Conflicts:
#   pyopencl/invoker.py
#   test/test_wrapper.py