Speed up workflow build caching

[strokovok]

• Remove default caching mechanism. Contents of ~/.cargo are now shared across runs since all runners are now on the same machine. Caching of target directory using remote storage is not efficient and can be implemented faster as written below.
• Wrap Rust compiler with sccache, which partly allows to reuse compiled dependencies (it's also mentioned here). Currently, there are 481mb of cache-hittable artifacts.
• Share target/debug and target/release across runs. This allows profits even on partial cache-hit from previous builds + saves a lot of disk space for us :)

[joshuajbouw]

Mmm, yeah conceptually this looks ok. Why the move away from cache@v2 though?

[strokovok]

Why the move away from cache@v2 though?

Well, cache@v2 works like this:

• In case of cache-hit it downloads ~700mb archive from GitHub storage and unzips it
• In case of cache-miss it builds from scratch, zips all artifacts (2.5GB -> 700mb) and uploads archive to GitHub storage
• Both of this steps work ~40secs in our case
• GitHub storage is limited to 5GB, so this cache-packages do not live long :)
• One change of dependency completely invalidates cache

There are multiple reasons to get rid of it:

1. Since we have our own machine with a lot of memory, we don't need to spend time for zipping/unzipping/uploading/downloading. We just can copy all the files as they are to some local cache in less than 1 second.
2. In case of cache@v2 it's only possible to work with huge monolith archives - one change of some dependency changes the hash and completely invalidates the whole cache. With our own machine we are able to use partial cache invalidation which significantly increases cache-hit probability for single dependency

[joshuajbouw]

Makes sense. Thanks for the explanation and the care and attention to that assessment.