Create a generic callbacks class for KubernetesPodOperator

[hussein-awala]

This PR adds a set of callbacks for KubernetesPodOperator called when the pod starts, terminates, and is cleaned up, and as when the client is initiated.

It provides some valuable params, like the pod spec, the Kubernetes client, and the current mode (sync or async); it could be helpful in different use cases like:

• create some resources based on the created pod and clean up them when the pod is deleted (for ex, a K8S service)
• patch some created resources to link them to the pod and make it the owner of these resources
• call an external system to trigger an event, push metrics...

The current implementation is only used in sync mode, but I provide the mode param for the future support for async mode. This support is blocked by the way of serializing the trigger kwargs, where all the provided types and methods are serialized as a string, and it's a bit difficult to recreate the original object without a custom serializer or impacting the event loop (for example provide the class path and import it in the triggerer)

[eladkal]

Also doc paragraph about this new cool ability is good to have

[hussein-awala]

@eladkal could you check the Google provider check? is it's complicated, I can bump the min cncf-kubernetes version to avoid this complexity.

[eladkal]

LGTM

[potiuk]

❤️ it

[potiuk]

There is a test failure though :(

[hussein-awala]

I just tested on a real cluster and found that the type I provide to the trigger is converted to a string. I even tried providing an object instance but had the same problem. That's because we don't serialize the trigger kwargs, but we keep sqlalchemy convert them to string (ExtendedJSON).

I need to revert the async part and keep it only for the sync part, and later we can add support for async callbacks once we find a clean solution for the issue; I will take this into account while working on the kwargs encryption.