[CI] Properly handle sccache file permission in Linux GPU workflow

[qiao-bo]

At the moment, we tar and copy the checked-out code from our GPU Linux job into the executed docker container.

taichi/.github/workflows/testing.yml

Line 273 in 80f20f2

tar -cf - ../${{ github.event.repository.name }} --mode u=+rwx,g=+rwx,o=+rwx --owner 1000 --group 1000 | docker cp - taichi_build:/home/dev/

This was implemented some time ago for sccache handling (#3559). We need a better way to handle this because this step introduces two major problems in our CI.

1. The specified PID and GID must match the buildbot's execution user. Note that the value 1000 as specified, typically this value is assigned to the first user on e.g., Ubuntu. In another words, if a buildbot has two users, and we use the second one to execute the docker container. We will not match the permissions in the file system and some errors such as cannot write to metadata.tcb will be reported.
2. This leaves our GPU Linux CI dirty. As we implicitly rely on the checkout code of the Linux job, which on our self hosted runner is of location github-actions/_work/taichi/taichi. This folder is not deleted between different runs. The problem now is analogous to Windows workflow.

Welcome any suggestions!

[qiao-bo]

CC @ailzhang @frostming @rexwangcc

@lin-hitonami any idea of what are the alternatives to handle sccache?

[ailzhang]

cc: @feisuzhu fyi this might be an interesting issue for you - a great start to cleanup our docker related mess :P

[frostming]

maybe i will get the time to restart #4776, now it has some conflicts with the permissions on the self hosted runners

[feisuzhu]

Issue is irrelevant now.