Support custom SweepContainer image

[davidagold]

What

Currently, sweep tasks use a single image defined in the sweep experiment with the model definition and training script backed in. To scale to arbitrary model/training regimes, we need to allow users to run sweep tasks with their own image. The workflow looks like this:

Desired workflow:

1. Install bacon CLI and deploy a Bacon stack (this step only need to be done once)
2. In an experiment repository, define your architecture and training script. Build and push (to ECR)
3. Deploy a capacity provider
4. Use the CLI to trigger a sweep_dag run, where the dag_config contains:

Acceptance Criteria

• Export a shortcut to deploy sweep task definition
• sweep_task is parametrized
• Sweep experiment config accepts task definition as argument

[davidagold]

Note that we'll need to modify the init_sweep operator to pass a parametrized project name.