Kernel transform strategy based on iname tags

[inducer]

This is intended to help revive https://github.com/inducer/meshmode/pull/217. It's used in https://github.com/inducer/meshmode/pull/220.

I'm not sure this has any business being here, but I didn't have the stamina to give meshmode its own PyOpenCLArrayContext subclass again. Nonetheless, I think that's what we should do. @matthiasdiener, I'd be grateful for your help here.

That means:

• updating those imports across the tests and examples (in meshmode), and
• making the pytest fixture customizable as to what actx classes to use,
• Same goes for grudge and mirgecom, as they will likely grow their own kernel metadata at some point.
• This also makes the other transform strategies redundant, so those can really go. (and the other meshmode kernels should be updated to use this one)
• So PyOpenCLArrayContext.transform_loopy here should really just raise. To still be able to run tests for this package, introduce a subclass that has a no-op transform_loopy.
• To avoid a huge performance pitfall, it's important that the array context provides a useful error if no applicable transform strategy was found.

Draft because:

• [x] Needs https://github.com/inducer/loopy/pull/410
• [x] Point requirements.txt back to main
• [ ] Should really live in meshmode.

cc @thomasgibson @alexfikl

[inducer]

If you tried this and dismissed it: cfa98e7 fixed a big problem with the transform cache. (not sure how I missed that)

[inducer]

Superseded by https://github.com/inducer/meshmode/pull/236.