[WIP] Add pytest-based tests and remove nose tests.

[ketch]

Since nose is dead, this PR begins the switch to using pytest in PyClaw. Some notes:

• I'm not using the intricate machinery from pyclaw.utils that helped to generate many variants of the same test. That code doesn't immediately work with pytest (which doesn't support yielding tests) and it seems in some ways preferable to have more straightforward (easy to understand) testing code, even if it means being a bit more verbose about the definition of the tests.
• In addition to the regression tests (which check that the output hasn't changed from what it used to be), I've added some accuracy tests where we run on a reasonably fine grid and check against the exact solution.

[mandli]

This works well for me.

I am onboard with the idea of making the tests more verbose and also adding accuracy tests where we can. What would be even better would be convergence but I imagine that being a bit tricky.

[rjleveque]

I also agree it would be best to make the tests more verbose and easier to understand and modify.

I wonder about running convergence tests, since part of the point of the tests is to make them run very quickly so that we do it frequently, and so they can be done automatically for PRs. Testing convergence is of course important, but once it's been determined that a particular test behaves properly in this regard, isn't it enough to test a few things coming out of one quick run to make sure the example hasn't broken in some way due to changes elsewhere? I thought that was the point of these unit tests, not to validate the numerical method.

We should talk about how to make similar changes in other repos, e.g. updating clawpack/amrclaw#279.

A couple other thoughts regarding the Fortran code tests:

• Rather than redirecting output to a scratch directory, I think it would be simpler if we just sent it to _output as usual. These tests all produce very minimal output, and when things go wrong it would be easier to find the output in the standard location.
• In the old system, when tests pass, the output directory is deleted. If possible it would be nice to add an option so that the tests can be run with the output retained, which is often useful if trying to do a dual debug to figure what went wrong in a new version where the test does not pass. Sometimes one can just do make .output at the command line to make the output, but sometimes the tests set up some data that is used in the test and then it's not so easy to figure out how to run the test and view the output (or at least that's my memory, admittedly I haven't done this in a while.)

[ketch]

Thanks for the comments! In the meantime, I did figure out how to setup similar tests with different function arguments in a more streamlined way: https://docs.pytest.org/en/7.1.x/example/parametrize.html

But maybe it's better to just be verbose and explicit. We can discuss it later today.