Fix numpy initialization. Closes gh-214

See merge request !18

Pytest is a test-requirement not a runtime requirement

Installing pycuda shouldn't bring in a test runtime, only
when you are planning to run the test suite of pycuda should
this be brought in.

Try to discover CUDA binary path, to add to Py3.8's DLL path (Closes gh-213)

See merge request inducer/pycuda!17

Fix a few left-over 32-bit-only memory sizes

See merge request inducer/pycuda!16

Update tools.py

Starting from compute capability 3.0, up to 64 warps can fit on a SM.

CI: Stop targeting specific Py3 versions

See merge request inducer/pycuda!15

allow other threads to run during device init() and make_context() since those...

See merge request !14

allow other threads to run during device init() and make_context() since those calls can take hundreds of milliseconds

Allow setup_requires numpy with updated aksetup

See merge request inducer/pycuda!13

Bump CIs to Py3.6, add Titan V job

See merge request inducer/pycuda!12

replaced deprecated numpy.random.random_integers with randint