Setup sccache for linux test (#84916)
The TTS alarm in HUD is quite noisy because of `pull / linux-focal-py3.7-gcc7 / test (backwards_compat)`. Some runs take up to 50m, i.e. [4893945786](https://github.com/pytorch/pytorch/actions/runs/3038960118/jobs/4893945786) while others take only 10m, i.e. [4893781147](https://github.com/pytorch/pytorch/actions/runs/3038943635/jobs/4893781147). Looking closer into their logs, it turns out that the longer runs have a much higher rate of cache miss. For example, [4893945786](https://github.com/pytorch/pytorch/actions/runs/3038960118/jobs/4893945786)
```
Compile requests 6487
Compile requests executed 6224
Cache hits 4975
Cache hits (C/C++) 4975
Cache misses 1227
Cache misses (C/C++) 1227
Cache timeouts 0
Cache read errors 0
Forced recaches 0
Cache write errors 0
Compilation failures 9
Cache errors 13
Cache errors (C/C++) 13
Non-cacheable compilations 0
Non-cacheable calls 16
Non-compilation calls 247
Unsupported compiler calls 0
Average cache write 0.096 s
Average cache read miss 11.681 s
Average cache read hit 0.040 s
Failed distributed compilations 0
Non-cacheable reasons:
multiple input files 15
unknown source language 1
Cache location S3, bucket: Bucket(name=ossci-compiler-cache-circleci-v2, base_url=http://ossci-compiler-cache-circleci-v2.s3.amazonaws.com/)
```
In https://github.com/pytorch/pytorch/pull/82103, we didn't setup `SCCACHE_S3_KEY_PREFIX` for `_linux-test`, which could explain the high rate of cache miss here. `backwards_compat` is a bit different than other tests in which it compiles PyTorch and gets the benefit from caching.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84916
Approved by: https://github.com/seemethere