Make copies from host to device blocking

[maleadt]

Otherwise the following simple example fails:

# allocate CPU arrays
A = Int32[1]
R = zeros(Int32, 1)

# allocate GPU arrays
dA = CLArray(A)
dR = CLArray(R)

R .+= 1
# XXX: the asynchronous `CLArray(R)` only executes here,
#      resulting in dR containing the result _after_ the increment
dR .+= 1

# dR != R
# dR: [2]
# R:  [1]

It's very surprising to me that we haven't encountered this issue with our other back-ends. I assume that CUDA.jl's handling of unpinned memory (which it first copies synchronously into a pinned staging buffer) masks a lot of issues here.

Still, we'll probably have to make this change in other back-ends as well, because the memory could just as well have been fried before the copy finishes.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 64.03%. Comparing base (8aca9d7) to head (3a19188).
Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #243   +/-   ##
=======================================
  Coverage   64.03%   64.03%           
=======================================
  Files           9        9           
  Lines         456      456           
=======================================
  Hits          292      292           
  Misses        164      164           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.