[testing] Tests mark themselves UNSUPPORTED if features aren't available

[rolfmorel]

Also loosens dependency upon uv for running tests -- all that is required is that lit is running in a suitable environment (e.g. through uv run lit . or by first entering an venv and then lit .).

Also splits lighthouse.utils into two so that the examples/tests that don't actually depend on torch accidently end up with that dependency (there was one such such instance already).

[rolfmorel]

E.g. if you didn't install torch dependencies (e.g. did uv sync without --extra ingress-torch-cpu flag), then we get:

$ lit .
-- Testing: 8 tests, 8 workers --
UNSUPPORTED: Lighthouse test suite :: examples/ingress/torch/mlp_from_model.py (1 of 8)
UNSUPPORTED: Lighthouse test suite :: examples/ingress/torch/mlp_from_file.py (2 of 8)
UNSUPPORTED: Lighthouse test suite :: examples/xegpu_matmul/matmul.py (3 of 8)
UNSUPPORTED: Lighthouse test suite :: examples/mlir/compile_and_run.py (4 of 8)
UNSUPPORTED: Lighthouse test suite :: examples/ingress/convert-kernel-bench-to-mlir.py (5 of 8)
UNSUPPORTED: Lighthouse test suite :: examples/llama/test_llama3.py (6 of 8)
PASS: Lighthouse test suite :: examples/schedule/transform_a_payload_according_to_a_schedule.py (7 of 8)
PASS: Lighthouse test suite :: examples/ingress/mlir_gen/generate-linalg-3layer-mlp.sh (8 of 8)

And if you then install torch but still forgot to/didn't fetch the KernelBench third_party repo:

$ lit .
-- Testing: 8 tests, 8 workers --
UNSUPPORTED: Lighthouse test suite :: examples/ingress/convert-kernel-bench-to-mlir.py (1 of 8)
PASS: Lighthouse test suite :: examples/schedule/transform_a_payload_according_to_a_schedule.py (2 of 8)
PASS: Lighthouse test suite :: examples/ingress/mlir_gen/generate-linalg-3layer-mlp.sh (3 of 8)
PASS: Lighthouse test suite :: examples/mlir/compile_and_run.py (4 of 8)
PASS: Lighthouse test suite :: examples/llama/test_llama3.py (5 of 8)
PASS: Lighthouse test suite :: examples/xegpu_matmul/matmul.py (6 of 8)
PASS: Lighthouse test suite :: examples/ingress/torch/mlp_from_file.py (7 of 8)
PASS: Lighthouse test suite :: examples/ingress/torch/mlp_from_model.py (8 of 8)

---

Without this change the above UNSUPPORTED tests would show FAILED instead.

[tkarna]

LGTM. At the moment we are only testing the "everything is installed" case in CI, right?

[rengolin]

why is the kernelbench still unsupported after installing torch?

[rolfmorel]

why is the kernelbench still unsupported after installing torch?

The example speaks for itself 😉

$ uv run examples/ingress/convert-kernel-bench-to-mlir.py
ERROR: KernelBench repo not found.
NOTE: Pull in dependency with: git submodule update --init third_party/KernelBench

[rolfmorel]

At the moment we are only testing the "everything is installed" case in CI, right?

Indeed. If people want this UNSUPPORTED mechanism tested in CI as well, do let me know.