image_tag, stdin experiments, CPU mode by default for TPU

- `--experiment_config` can now take experiment configs via stdin (pipes, yay!);
  specify `--experiment_config stdin`, or any-cased version of that, and the
  script will wait to accept your input.

  As an example, this command pipes in a config and also passes `--dry_run` to
  show the series of jobs that WILL be submitted when the `--dry_run` flag is
  removed:

  ```
  cat experiment.json | caliban cloud -e gpu --experiment_config stdin --dry_run trainer.train
  ```

  You could pipe the output of a nontrivial python script that generates a JSON
  list of dicts.

- `--image_tag` argument to `caliban cloud`; if you supply this it will bypass
  the Docker build and push steps and use this image directly. This is useful if
  you want to submit a job quickly without going through a no-op build and push,
  OR if you want to broadcast an experiment to some existing container.

- if you supply a `--tpu_spec` and DON'T supply an explicit `--gpu_spec`,
  caliban will default to CPU mode. `--gpu_spec` and `--nogpu` are still
  incompatible. You can use a GPU and TPU spec together without problems.

Change-Id: Ieea1467163f374bab01010c6a439dfff5877920f

cli hotfix.

Change-Id: I871b45f41a0a4a0369a1c73361bc37c3098a185f

move docs

Change-Id: Ib7183f332e5ad89b291f8fbe491a5b0099d74685

expansion bugfix

Change-Id: I7bedec45516fd289e73dd50f6aa9b364edc914f8

update changelog

Change-Id: I713d0d19cff2a4866919abbf1569a694b9bb0b3d

Experiment Config, GPU Spec and Machine Type

- `caliban cloud` now supports:
  - `--gpu_spec`, which you can use to configure the GPU count and type for your
    job.
  - `--machine_type` allows you to specify the machine type for all jobs that
    run.
  - `--experiment_config` lets you submit batches of jobs at once.
  - `--force` skips all validations and forces a submission with the specified
    config.
  - `--dry_run` will generate logs showing what WOULD happen if you submit a
    batch of jobs.

All of these validate as early as possible so that it's not possible
to attempt to submit a job that has obvious mismatches in GPU
type, count, machine type and region.

- The `caliban cloud --stream_logs` argument is now gone; the command prints,
  so this is easy enough to run without special help, and the argument made
  batch job submission difficult.

Change-Id: Ia0afcf5000390a6c50a9e8532e6994aaab2683dd

Add new flags, whole project --version flag

This PR:

- adds flags for region and project_id in cloud mode
- adds a robust "--version" option
- adds a flag that lets you specify a custom jupyterlab version
- removes the default project_id of MY project!
- updates the README with new options

Change-Id: Ia3db1f96ffecc1e3123402000e75f8e312b4256c