[CI] Refactor of tvm.testing.requires_* annotations

[Lunderberg]

This PR has two main goals.

• Improve the pytest messages when unit tests are skipped. When investigating the bug in [Hexagon]Use requires_hexagon instead of requires_hexagon_toolchain if running on hexagon target #11294, reporting the reason why a test is skipped (disabled in config.cmake, disabled in TVM_TEST_TARGETS, or no device available).
• Distinguish between tests that only require compile-time support from tests that require a physical device. The existing @tvm.testing.requires_* decorators maintain the same semantics, skipping a test unless both compile-time and run-time support are present. This adds an optional argument to indicate that the test does not require a physical device (e.g. to require cuda library support without requiring a cuda-capable GPU: @requires_cuda(support_required="compile-only")

cc @Mousius @areusch @driazati

[Lunderberg]

@mehrdadh I expect this to conflict with these changes in PR#11294. Between the two, your PR takes priority, as it is solving an immediate issue. If all goes well, I don't expect the conflict to be difficult to resolve, as the corresponding change after this refactor would be to replace target_kind_hardware="hexagon" with run_time_check=lambda: "ANDROID_SERIAL_NUMBER" in os.environ in the definition of requires_hexagon, and to replace requires_hexagon_toolchain with requires_hexagon(support_required="compile-only").

[mehrdadh]

Thanks for the refactor! LGTM!
cc @areusch and @driazati to take a look as well