[Enhancement] GPU RayCluster doesn't work on GKE Autopilot

[kevin85421]

Why are these changes needed?

GKE's Autopilot does not support GPU-using init containers, so we explicitly specify the resources for the init container instead of reusing the resources of the Ray container. The resource consumption of the init container should be constant. Hence, it is OK to hard-code the resources.

Related issue number

Closes #1349

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

# Step 1: Create a GKE Autopilot cluster
gcloud container clusters create-auto kuberay-gpu-cluster --region=us-west1

# Step 2: Install a KubeRay operator with this PR
# Step 3: Install a RayCluster https://gist.github.com/kevin85421/ef4c3a0c7493e76edd3d2cab5cea33e4
# (Note that you need to add nodeSelector)

[architkulkarni]

Nice fix! Once this makes it into the kuberay release, we can probably simplify some of the GKE instructions in our doc and just recommend users to use GKE autopilot going forward. (Depending of how stable we believe GKE Autopilot is)

[architkulkarni]

A comment here with some indication of why these numbers were chosen would be useful, perhaps a brief reminder of what the init container does

[kevin85421]

Updated 9039f24