CI: Reduce traffic and slow down size-bloating of `gh-pages`

[leofang]

Related to #2197.

I think we can probably pack all per-PR artifacts (doc preview + test report + coverage + …) to Read The Docs (RTD), and keep gh-pages lean and low traffic. We couldn't do this when cuda-python was not fully open sourced.

The idea I looked into in CuPy (cupy/cupy#10185) is

• take wheel from GHA and build docs in RTD, or
• build everything (wheels, docs, …) in GHA and take final artifacts to RTD
  I think RTD supports both, but I haven’t found time to set it up. Probably should prioritize for it and do it for both projects.