Consistent test_gpu flag

[ianthomas23]

Fixes #1130.

This simplifies the test_gpu flag so that it is boolean. Previously it could be True, False or None (meaning unspecified) and the None option was supposed to behave the same as False but with different test skipped messages. But it wasn't handled consistently, now it is. It still defaults to not running GPU tests. I am not sure this is the best default, but it can be dealt with in a later PR.

On a system with a GPU and with cupy and cudf installed, running the tests in either of these ways runs exactly the same tests (i.e. ignores the all GPU tests):

pytest datashader/tests
DATASHADER_TEST_GPU=0 pytest datashader/tests

and to run all tests including GPU ones use

DATASHADER_TEST_GPU=1 pytest datashader/tests

Also tested all 3 options in an environment without cupy and cudf installed, and if you set DATASHADER_TEST_GPU=1 then the test_gpu_dependencies test fails as expected.

After this PR, all tests run on CPU and GPU on my dev machine (Ubuntu 22.04 x86_64, Nvidia Quadro T1000) with the exception of that identified in issue #1131, which can be manually disabled using the -k "not log_axis_line" command-line argument.

[jbednar]

Looks good, thanks!