This PR improves build and link time for gcs_server_lib from 7m49.451s to 6m55.810s on my devbox (>10% improvement).

In this PR, I tried to do a few things:

• Split the giant bazel target into smaller pieces, which better leverages bazel's parallelism
  • Other benefits see [core] De-glob bazel dependencies #50243 (comment)
  • I'm doing it half-manually, better to leverage tools if possible
• Leverage forward declaration if possible, so next time header change won't trigger rebuild for unnecessary dependents
  • This is also the benefit for smaller bazel targets
• Place implementation into impl file rather than header file
  • Similar to above, also useful to reduce re-compilation
• Remove unnecessary header and dependencies
• Split dependencies into smaller granularity, apply to a lot of targets under ray/util and ray/common folder

The build time is still intolerably long, eg, a state_util target takes >120 seconds...
I've worked on large C++ binaries, which exceed linker 4GiB linker limit, which have to leverage some hacky ways to work around; but the compilation/link time is still under 15 minutes.

Toolchain setup wise, build file auto-generation and IWYU might be helpful;
but I personally feel it's more easy and important to reduce the interdependency between targets.

Benchmark command for reference:

• bazel clean
  • cleanup local bazel cache before benchmark
• sudo sync; sudo sh -c "echo 3 > /proc/sys/vm/drop_caches"
  • deprecate filesystem page cache, which I found affects performance by >10%
• bazel build //:gcs_server_lib