[Docker] - Refactor and automatic upload of containers to repo

[goliaro]

Description of changes:

This PR refactors the Docker containers, to make the builds more efficient, and enables the user to pass build configs through the use of environment variables (passing build configs to Docker got partially broken after #392 ). In addition, we add a mechanism to automatically publish the environment and flexflow containers to ghrc.io. The CUDA version of the uploaded docker image is compatible with any CUDA GPU architecture. This gives us several advantages:

• users can just pull the flexflow docker image if they want to try flexflow under the default configs
• CI tests can use the flexflow-environment image to avoid having to re-install all FlexFlow dependencies in each test

Related Issues:

Linked Issues:

• Issue #

Issues closed by this PR:

• Closes #

Before merging:

• Did you update the flexflow-third-party repo, if modifying any of the Cmake files, the build configs, or the submodules?

[lockshaw]

Overall LGTM.

If the bash scripts in docker/ continue to get more complicated, we should really consider porting them over to python. Just having argparse, etc. would probably already reduce the amount of code significantly, and remove a bunch of more complex and unusual bash features.

[goliaro]

Overall LGTM.

If the bash scripts in docker/ continue to get more complicated, we should really consider porting them over to python. Just having argparse, etc. would probably already reduce the amount of code significantly, and remove a bunch of more complex and unusual bash features.

yes, good point. Perhaps we can switch to Python when adding the pytorch align tests in CI