Break down CTest cases into each test suite

[ISSOtm]

This allows parallelising the individual programs' test suites, which seems to save around two minutes and a half!

This is a medium-effort preview of #1927: this doesn't expose each test to CTest, but at least we get some useful amount of parallelism out of this. (Especially for our CI.)

[ISSOtm]

This seems to save a couple of minutes on Windows, but we're still in the realm of 10 minutes for testing. Though at least CTest is nice enough to give us a (wall-clock) breakdown:

1/6 Test #6: fetch-test-deps ..................   Passed    8.47 sec
2/6 Test #4: rgbgfx ...........................   Passed   29.16 sec
3/6 Test #2: rgblink ..........................   Passed   36.93 sec
4/6 Test #3: rgbfix ...........................   Passed  103.54 sec
5/6 Test #1: rgbasm ...........................   Passed  230.78 sec
6/6 Test #5: external .........................   Passed  357.14 sec

...which seems to imply that the slowest is running the external tests. For comparison, macOS (5m30s):

1/6 Test #6: fetch-test-deps ..................   Passed    3.66 sec
2/6 Test #2: rgblink ..........................   Passed   42.45 sec
3/6 Test #3: rgbfix ...........................   Passed   66.79 sec
4/6 Test #4: rgbgfx ...........................   Passed   67.17 sec
5/6 Test #1: rgbasm ...........................   Passed  179.85 sec
6/6 Test #5: external .........................   Passed  155.11 sec

[Rangi42]

Each external test is independent; could they at least be broken down into free and nonfree categories?

[ISSOtm]

That sure would be nice! But the test runner currently doesn't allow splitting these two categories. And I've limited my effort in this PR to not touching the test runners at all, as a scope limiter.

[Rangi42]

If we're able to run ctest before cmake --install, then we don't need to be installing rgbds at all. That goes for the other OSes as well.

[ISSOtm]

I think installing is mostly used as a control for whether it works and in the correct locations. But, yeah, at this point we should probably just get rid of all --installs—in a follow-up PR, though.

[Rangi42]

There were some comments to follow up on in more PRs (getting rid of unnecessary install commands in CI, using ProcessorCount instead of hard-coded 4, consistent ctest flags instead of --label-exclude external-repos being unique) but this is good enough. Thank you!