Improve CI/CD

[nepalez]

In this PR I tried to improve the CI in several aspects. Below is the description of commits in this PR:

1. Turn off incremental compilation

The incremental build is a way to fasten future compilations for the cost of extra dependency tracking and bigger /target.
This approach is not necessary in the CI because every compilation is done from scratch.

As a result of turning it off the size of the target reduces from 660MB to 543MB.

2. Turn off debug info

This change reduces the size of the /target directory from 543MB to 424MB without (or from 660MB to 483MB with) the incremental compilation.

3. Improved artifacts caching

3.1. Use Swatinem/rust-cache@v2 to make caching smarter.

The primary difference is that this GHA caches not the whole content of the ./target but only dependencies. Along with points 1 and 2 this makes caches 3x smaller (230-450MB instead of 0.8-1.3GB) for the same build time.

3.2. Split build and test stages

This enables to store the cache if the build succeeded even though the test can later fail. It also useful to immediately see the stage where a failure occur.

3.2. Use the shared_key in caches instead of the job-based key

This enables some caches (primarily ubuntu/stable) to be reused between jobs. It affects fast jobs only (like checking examples, running clippy etc., which are now reuse the caches built by test matrix), but anyway this approach consumes less resources.

3.3. Use v4 (instead of v3) of the actions/setup-go

This version provides caching by default (104MB x 3 = 312MB)

3.4. Drop unnecessary Go/nats-server installation for checking msrv and examples

These steps weren't necessary because we only run rust check without even preparing a full build, not to mention running tests.

3.5. Add the cache for the cargo-spellcheck

It saves ~4 min at the corresponding job for the cost of the additional 54MB cache.

3.6. Add the --no-deps option to cargo doc

I thought it's up to the dependencies' maintainers to check their docs

---

What I would also add is hashing prepared Cargo.lock instead of the **/Cargo.toml.

While this can reduce the lifetime of caches, we could see possible problems with the latest versions of dependencies.

This change would makes the msrv check slower (building lockfile on 1.67.0 takes ~2-3 min), but it wouldn't be critical (this job is 2-3 times faster than running the test suite).

I made this change in a separate commit, but would like to discuss it before adding the commit to this PR.

[levb]

LGTM, but I am not really a Rust guy, don't know enough about the specific actions used here; Also @Jarema, I don't really understand the Rust caching, so please check that the cache key makes sense to you.

[levb]

1/5 we should probably use the latest server release rather than main for benchmarking

[levb]

curl -s https://api.github.com/repos/nats-io/nats-server/releases/latest | jq -r '.tag_name'

[nepalez]

After I added the caching here, this would require getting the current version of the server to use it in the cache key (for now I only build the key from the os and the go version)

[nepalez]

I also updated the last commit with caching the nats-server binary as well. This saves ~3min on windows (and only ~1 min on linux/macos). The additional size of 3 caches is 3x9MB = 18MB, so it looks like a good deal.

[Jarema]

LGTM!

Thanks for your contribution!