Fix running under Docker following 4.0.0rc4 config changes

[ebr]

Summary

• A new container without an existing root directory would fail to start due to permission issues and being unable to create the config file - this is fixed.

• The --host CLI argument was removed as part of config optimization, and we were using it in the Dockerfile. Because the default bind host is 127.0.0.1, the app would not be accessible from outside the container (even from local host). Anyway, this PR adds INVOKEAI_HOST and INVOKEAI_PORT to the Dockerfile to provide working defaults that do not require additional intervention from the user.

A couple of other improvements to the dockerized experience:

• INVOKEAI_ROOT can now be set to a relative path without using a dedicated env var pair for this purpose. The original (HOST_ and CONTAINER_) root dir environment variables are retained because we know some users have these special use cases. The env.sample file now explains their use more thoroughly.
• No longer pre-install Torch in a separate Docker layer. Thich reduces the image size and removes the need to manually bump Torch/torchvision versions in the Dockerfile - we now just let pip do its thing. We do however still need to maintain the extra-index-url logic in the Dockerfile to ensure cross-GPU compatibility.

QA Instructions

• on Linux (CUDA and/or ROCm), enter the docker dir, and either execute run.sh or docker compose up - this will run with all defaults.
• terminate the container (docker compose down), then try to set INVOKEAI_ROOT to a new path on the filesystem, and run it again. In all cases, the container should start and the new 4.0 UI should be usable as expected.

Merge Plan

Merge at will

Checklist

• The PR has a short but descriptive title
• Tests added / updated - N/A
• Documentation added / updated