[ci] Improve docker image for Linux CPU build and test

[qiao-bo]

Currently, we employ an Ubuntu 18.04 based docker image (Dockerfile here) for building and testing Taichi's CPU backend. The current image is approximately 12GB, we would like improve this image by reducing the image size to its minimal.

Things to consider include:

• Familiar with Taichi's build environment and workflows.
• Optimize Dockerfile by following some of the best practices in building (For example enable multi-stage build and minimize number of layers, etc) and making use of optimization tools such as Dive and Docker-slim.
• Ideally we want to reduce the image by 50%, and being tested on Ubuntu and Arch linux.

After exploring the previous steps, we can also move on to optimize other backends:

• Improve docker image for GPU build and test
• Improve docker image for CPU build and test (Windows)
• Combine the multi-stage build approach with our generator scripts.

[frostming]

If we switch to the container setting in workflow(#4776 ), we can get rid of the conda setup since it can be done by a Github Action: setup the demanded python version rather than pre-installing a bunch of python versions

[qiao-bo]

If we switch to the container setting in workflow(#4776 ), we can get rid of the conda setup since it can be done by a Github Action

that would be great! conda environment occupies quite some space.