nsakabe-fixstars

Refer to https://github.com/cupy/cupy/pull/78

The case failed 5 times in 100 repeats.

Hypothesis: [sequential clrandGenerator::_issue_by_np](https://github.com/fixstars/clpy/blob/113daf35e92af558e032fb8daa35382d3eafff2c/clpy/backend/opencl/random.pyx#L103-L106) can be interleaved among threads

@ykitta-fixstars Could you handle this?

Summaries of difference between (CuPy) versions: https://docs-cupy.chainer.org/en/stable/upgrade.html

> We tried to implement an equivalent for `cupy.cuda.PinnedMemory` and found that there isn't correct substitution in OpenCL of allocating pinned host buffer. https://github.com/fixstars/clpy/pull/144#issuecomment-463496350 The working branch was [6-alloc_host_ptr](https://github.com/fixstars/clpy/tree/6-alloc_host_ptr).

64-bit atomic operations need `cl_khr_int64_base_atomics` extension. `clinfo` says NVIDIA driver doesn't implement this. And I couldn't find how to implement 64-bit `atomicAdd`s without it...

I'm working on this issue with the policy of supporting only 32-bit operations.

**Note for developers:** When you implement 64-bit integer `atomicAdd`, please test it by adding `numpy.uint64` into [test_ndarray_scatter.py](https://github.com/fixstars/clpy/blob/46b9a6f60b5732bf5167929bf88b75ed5332ebe6/tests/clpy_tests/core_tests/test_ndarray_scatter.py) test target.

Yeah. Revert bbdec43 and remove corresponding TODO notes.