tests should be able to run in every pr

[sozercan]

currently, tests require approval. since this is a CNCF project, this is a barrier for entry for new contributors and decrease pr merge velocity.

[zhuangqh]

We added this restriction to e2e test to prevent malicious users from consuming excessive GPU resources. Do you have some suggestions about this use case?

For now, Every step needs an approval if it's contributor's first PR. After that, unit-test and other general test would be triggered automatically. Still need an approval to run e2e test.

[sozercan]

I think we should be able to test it on github agents using kind with CPUs, and have separate GPU workflows that test with manual approval from maintainers when we think it needs to be tested (like specific PRs that modify provisioning, before releases, etc)

There should be no credentials or permissions required for the general PR e2e flow, so contributors can create a PR without waiting for an approval.

I can take a stab at this. Can you assign to me @zhuangqh

[sozercan]

this is also tied in with #984 so we can use quantized models with cpu for quick tests