There is no reason to run simd_op_check so much

[alexreinking]

For example, we run simd_op_check_arm for every HL_JIT_TARGET configuration we test on x86 Windows/macOS/Linux. Unless we're on ARM, where differences in the host capabilities affect simd_op_check's ability to run the tests, we need to test this at most once.

Some of these tests are minutes long individually. For one data point: simd_op_check_hvx runs for 2 minutes on x86-64-osx. But then that builder does it eight times. 16 minutes spent re-running that test. And that's just HVX!

[alexreinking]

One possible resolution: the tests can SKIP themselves unless they're running on HL_JIT_TARGET=host. Slight adjustments needed for wasm and hvx.

[mcourteaux]

I think we do want this guard ONLY if we're running the CI (buildbots / GHA). Local tests on a single developers machine should not skip those, in case we want to run the test suite locally. So testing against HL_JIT_TARGET=host is only fine on CI runs.

[abadams]

On linux at least the problem is pointlessly running it under different HL_TARGET settings. Maybe we could skip the test if on CI and target is anything other than host?

Edit: oops, that's exactly what you already proposed!