feat: make building the docker image configurable.

[luc-hiverge]

This commit adds a new top-level section to the experiment config:

docker:
  dockerfile: exp1/Dockerfile
  verbose: true

which enables the user to pick the Dockerfile for the baseimage of the sandbox. It also allows making the build-process verbose, i.e. --process=plain.

[luc-hiverge]

This allows me to create an "overlay" style setup:

$ ls
checkpoints  experiments  snapshots  src   tools

where the source code to be optimized lives in src (it's "almost" unmodified upstream fairchem).

Then I can create different experiments like this:

$ tree experiments
experiments
├── uma
│   ├── Dockerfile
│   ├── evaluate.py
│   ├── hive.yaml
│   ├── README.md
│   ├── run_benchmark.sh
│   └── uma-speed.yaml
└── wigner
    ├── Dockerfile
    ├── evaluate.py
    ├── hive.yaml
    └── README.md

Then the dockerfiles do:

COPY src /app/repo
COPY experiment/uma/... /app/repo

(the overlay system).

Then I launch experiments as follows:

$ cat experiements/wigner/hive.yaml
project_name: wigner
coordinator_config_name: lucg-uma
docker:
  dockerfile: experiments/wigner/Dockerfile
runtime:
  num_agents: 5
  ...
$ hive create exp uma-wigner- -f experiments/wigner/hive.yaml 

[brp-hiverge]

/lgtm

[kerry-hiverge]

/kind feature

[kerry-hiverge]

/approve