Refactor CI workflows and improve test configuration

[alexreinking]

This PR bundles a few changes together (to avoid spending a bunch of CI time these changes are all designed to reduce)

• We skip simd_op_check builds on non-host targets
• We move scattered autoscheduler tests to tests/ (this might be slightly broken; we'll see what CI says. I tested it to work locally, though)
• The GHA workflow steps have been grouped by path-relevance. That way we don't run tests that couldn't possibly fail for a code-related reason.

Fixes #9009
Fixes #9004

[mcourteaux]

If every runner targets host, then every runner will simd op check all targets ? Wasn't the goal to not do it redundantly ?

This could be stripped further down to every machine just testing its own architectures op check suite.

[alexreinking]

I figured we wanted to make sure cross-compilation worked. But maybe we have enough coverage elsewhere. I'd like @abadams to weigh in

[shoaibkamil]

Looks good to me, except not sure about change in one of the autoscheduler sources. See my comment

[alexreinking]

win-worker-3 is timing out on halide_as_onnx_backend_test (which is very heavy). Will investigate in a follow-up PR.