Implement UpdateModel backend workflow

[petermuller]

Creates UpdateModel state machine implementation with frontend handler validations to force as many synchronous validations as possible before actually hitting the async workflow, which is a much user friendly experience, and easier to debug.

Model only polls for capacity between stopped and in-service states as the model is still functional in other scenarios. When starting a model, the state machine first waits for the desired number of instances to spin up, THEN it waits for the user-defined warmup time before adding the model to LiteLLM, that way customers can't try to make inference requires prior to the models actually spinning up. the model instances pop up healthy before the models are fully initialized, so this wait is necessary to ensure that the models are working before we open them back up to requests.

known issues:

• UI does not seem to send the correct payload for updating ASG configuration to the backend
  • UI is sending autoScalingConfig instead of autoScalingUpdateConfig
• UI cannot update from embedding to textgen or from non-streaming to streaming
• UI does not allow a user to start a stopped model without refreshing the page first if the model was stopped in that browser session

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.