[CI] Generate manylinux2014-compatible wheels with CUDA backend in release workflow

[strongoier]

Related issue = close #4395

The docker image produced by #4491 was uploaded as registry.taichigraphics.com/taichidev-manylinux2014-cuda:v0.0.0. Then this PR adds a job in release workflow to build and test the new wheel with CPU and CUDA backends in that image.

The new wheel is named as taichi-0.9.2-cp38-cp38-manylinux2014_x86_64.whl, while the original one (with OpenGL/Vulkan support as well) is renamed as taichi-0.9.2-cp38-cp38-manylinux_2_27_x86_64.whl (the original name with manylinux1 is wrong). pip install will choose the wheel closest to the user platform.

[netlify]

✔️ Deploy Preview for docsite-preview canceled.

🔨 Explore the source changes: 8316cb6

🔍 Inspect the deploy log: https://app.netlify.com/sites/docsite-preview/deploys/6231a189eb0f530008940e2c

[frostming]

Mostly LGTM but is there any chance to merge it into the matrix of build and test linux?

I see the only difference is the image name and TI_WANTED_ARCHS(does it make difference if we make them the same for this env var?)

[strongoier]

Mostly LGTM but is there any chance to merge it into the matrix of build and test linux?

I see the only difference is the image name and TI_WANTED_ARCHS(does it make difference if we make them the same for this env var?)

I agree that there are some similarities, but there are still other differences like:

1. Different TAICHI_CMAKE_ARGS
2. No -v /tmp/.X11-unix:/tmp/.X11-unix
3. No -e DISPLAY

I am not sure if putting these together will make things clearer.

[frostming]

@strongoier Ok then that is fine to keep it as it is now. I didn't compare them carefully.

[ailzhang]

LGTM!

[qiao-bo]

LGTM! In the future, we might need to think about adding to test.yml as well.