VT Proto migration

[gavins-db]

Migrate from GoGo to VT proto

• I added CHANGELOG entry for this change.
• Change is not relevant to the end user.

Changes

• Migrate thanos from gogo/protobuf to vtproto with pointer-based labels ([]*Label and other types). This is the majority the changes.
• A custom gRPC codegen plugin for pool-aware unmarshal.
• VT protos also have label string interning during unmarshal which is a big reason for the memory improvements in this branch. A custom pkg/unique string interning package (xsync.Map + finalizer cleanup, with an optional no-GC variant via build tag) improves performance here compared with stdlib unique which itself was already an improvement on the previous string interning library.
• Together these yield ~50% reduction in B/op and ~40% reduction in allocs/op with a minimal cpu increase on the gRPC receive synthetic benchmark. - This cpu increase may actually be the in the nature of the synthetic benchmark and performance may (I hope anyway) be better in a real-life scenario.
• note: cortex code has not been touched.

This PR was primarily focused on optimizing the receive path. It is possible there are some regressions in other areas (e.g. Query) however, at this point we should just see if this is a viable way forward at all. If it isn't then we stop here. I can close this and follow up with some of the micro optimizations i've made (or planned to make) in a separate PR building on gogo. If this is a viable path forward then I/we can begin tracing down further regressions and making improvements such as enabling pooling in other places. and even re-thinking how the underlying TSDB works in prometheus to carry these changes all the way through -- which I believe could have a greater impact.

Verification

• existing tests pass.
• introduced additional benchmarks which I've used to check on things.
• only a real test will tell...

[gavins-db]

🔁

[gavins-db]

reopening to force tests to re-run 😈