Updating failed APIs is slow

[ospillinger]

To Reproduce

1. Deploy an API with a handler bug on a single GPU machine with a 1 GPU resource request
2. Wait for the failure status
3. Fix the bug and redeploy

Actual Behavior

Failed API takes a long time to terminate and the compute resources are not available for the updated API.

Expected Behavior

Updated API deploys quickly.