Avoid repeating target-independent tests

[alexreinking]

We currently run many JIT / correctness tests for GPU targets even when those tests do not have GPU schedules. For instance, it does not make sense to run correctness_custom_allocator once per target since it's CPU-only. There are very many tests like this.

[abadams]

It probably wouldn't be very hard to just make more groups in test/correctness/CMakeLists.txt instead of just "correctness". E.g. we could have a CPU group, and GPU group (things that skip if no gpu target is enabled) and a "both" group. Classifying tests into groups initially could probably be done with a grep invocation. When new tests are added the human that adds them can decide which group it should go in.

[alexreinking]

I agree: this could be done at the CTest/CMake level fairly easily (ie. no source changes). It would just be a touch tedious to validate that everything was sorted correctly.

[abadams]

As a first cut, anything without the string "gpu" in it should not be run when testing GPU backends, and anything without the string "hexagon" can be skipped when testing hexagon.