Instead of building once on a CPU runner and once on a GPU runner, we can build once on a runner that allows both (basically a GPU runner) and upload the test artifact with different feature flags for both CPU and GPU test runs.

Currently all tests panic on this in OVM:

/// Returns the given cargo Package from the metadata in the Cargo.toml manifest
/// within the provided `manifest_dir`.
pub fn get_package(manifest_dir: impl AsRef<Path>) -> Package {
    let manifest_path = manifest_dir
        .as_ref()
        .join("Cargo.toml")
        .canonicalize()
        .unwrap();

I think this error might happen because manifest_dir is different between our self-hosted GPU server and the warp runners. Previously both CPU build and CPU test run happen on warp runner, but now CPU build happens on self-hosted server while CPU test run on warp runner.

Probably need to wait for setting up some warp runners to test this again.