Register sccache epilogue before starting sccache (#92587)
Fixing XLA test job flaky with sccache failing to start with a timeout error, for example:
* https://github.com/pytorch/pytorch/actions/runs/3953719143/jobs/6770489428
* https://github.com/pytorch/pytorch/actions/runs/3952860712/jobs/6769339620
* https://github.com/pytorch/pytorch/actions/runs/3946315315/jobs/6754126326
XLA test job actually builds XLA as part of the test ~~, so it needs sccache~~
* Register sccache epilogue before starting sccache, so that any errors when starting sccache can be printed
* Add `-e SKIP_SCCACHE_INITIALIZATION=1` to `_linux_test` workflow, this is the same flag used in `_linux_build` workflow. Quoted the reason from the build script:
> sccache --start-server seems to hang forever on self hosted runners for GHA so let's just go ahead and skip the --start-server altogether since it seems as though sccache still gets used even when the sscache server isn't started explicitly
* Also fix the code alignment in `.jenkins/pytorch/common-build.sh`
* We don't even use sccache in XLA test job, but there is an S3 cache used by bazel there (`XLA_CLANG_CACHE_S3_BUCKET_NAME=ossci-compiler-clang-cache-circleci-xla`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92587
Approved by: https://github.com/malfet, https://github.com/ZainRizvi
