Feature/loop for crashed tasks

[rakovskij-stanislav]

Karton already supports debug mode, but sometimes you need to works with already "unwealthy" tasks that crashed.

Restarting them is not an option - you cannot predict which consumer will get this task, while you need to feed this task to exact consumer you run under debugger / with extra prints / etc. The goal is to debug crashed task without interrupting the processing.

This PR introduces new method, Consumer.loop_crached_tasks that gradually makes debugging process more easy by working only with crashed tasks. It has several restrictions:

• It ignores timeouts (for benefit of debugging).
• It works with karton.tasks instead of karton.queue list in redis.
• It does not register new binds, shut down already running instances.
• It does not increment CRASHED metrics on further crashes.
• It does not immediately change this task status to STARTED.

[rakovskij-stanislav]

@psrok1 Hello! Just a gentle ping regarding this PR)

[psrok1]

Hello! Thanks for the contribution and sorry for very late answer. We had some internal discussions about this PR and then it was put on the shelf.

I don't think that this method should be a part of Consumer class if it has so many constraints:

        - It does not rely on `karton.queue`. It finds crashed doc in `karton.task`.
          So RUN ONLY ONE REPLICA to avoid race condition
          and large resource consumption.
        - It does not rely in task_timeout.
        - It does not register new binds.
        - It does not shut down another instances on binds / version mismatch.
        - It does not listen queue in a traditional way / it dows not subsribe.
          It looks for tasks in `CRASHED` state.
        - It does not increment `TASK_CRASHED` metrics.
        - It reimplements `Cunsumer.internal_process` in simplified way.

If we add something to one of the main Karton classes, we also set a contract that we will not break it in future versions. This feature looks like something really difficult to maintain and it's possible that we're going to add much more points to that list in the future 😃

We thought about implementing it in a different way, but right now we don't have any ideas. I would make an issue on that problem and maybe someone will come with less hacky way to implement such utility.