Lazy initialization of input readers and evals

[droot]

Two separate but slightly related changes (captured in separate commits)

Lazy initializations of Input readers (tty and readline)

Initialize these readers only when needed. Two benefits:

1. We won't setup these when running in non-interactive mode (--quiet)
2. k8s-bench feeds the interactive script on stdin, that forces the kubectl-ai to use tty for input when the input is through stdin. And github actions don't make /dev/tty available and that causes k8s-bench to fail when running in CI (github actions)

CI changes

1. Added step to run lightweight evals (two tasks related to scaling)
2. Added a workflow for periodic evaluation

Note: I am using keyless setup for authenticating github actions with GCP's vertexai platform. I expect a few bumps. I have tested this in my own fork, but let see how it goes here.