Document image metadata and etcd data structures

Let's document an old issue which is implemented in many ways such as configuration APIs and enhanced CLI supports.

- Let's remove the necessity of managing etcd manually. Just pulling images in agents should be sufficient for users.\* When scanning install images in the agent, let's scan metadata labels as well.
- The following metadata will be used:\* `ai.backend.limits.*.min`, `ai.backend.limits.*.preferred`, `ai.backend.limits.*.soft`
- `min` is the minimum required amount.
- `preferred` is the default amount.
- There is no maximum limit: `preferred` is the default maximum limit and a scaling-group configuration can override it. (`0` means no limit except agent's host limits)

- Automatically generate aliases (while keeping manual aliases):\* `labup/kernel-python-tensorflow:1.12-py36-cuda10`=> (`tensorflow`, `python-tensorflow`) x (`:1.12-cuda`, `:1.12-py36-cuda`, `:1.12-cuda10`, `:1.12-py36-cuda10`)
- If there are multiple platform tag versions and the kernel creation request does not specify a specific version, prefer latest ones.
- The accelerator tags are always required to use with accelerators of the same type.

JIRA Issue: BA-327