Try out offficial `prometheus_client` for metrics

[clux]

A working POC for prometheus/client_rust ripping out tikv/rust-prometheus. Not much easier than measured (via #71) because the documentation is lackluster, but the codebase ends up being slightly nicer (and proc macro stuff less confusing for sure).

PROS:

• similar conventions to what everyone is used to from the tikv crate
• Family<ErrorLabels, Counter> style generics works well and is easy to setup
• exemplar support present and working (using it for reconcile duration histogram)
• Registry::with_prefix that allows sharing of metrics modules
• Unit support in help strings nice minor thing

CONS:

• barebones, one real example in docs
• necessary manual arc wrapping of registries - was undocumented and had to apply measured's patterns manually
• minimal protobuf support out-of-date >7mo, barely tested (all examples use text encoder), undocumented in crate docs, doesn't integrate with frameworks as prost is out of date (e.g. actix-protobuf has prost@0.12)
• 1 documented exported encoder returns a String
• benchmarks from measured makes this look awful in comparison despite their goals
• also minimally maintained: https://github.com/prometheus/client_rust/issues - not even merging passing dependabot prs
• nit: not really qualify a client.

This is not updated well enough that I would be comfortable using it. If it had working protobuf and maybe native histograms or something this would be worth talking about, but it's only really bringing some slight ergonomics + exemplars to the table, that's useless if they don't update depedendencies. We don't have to choose this. There's no major downside for using an unofficial metrics crate if offical upstream is just a collection of unoptimized structs putting client in its name to make itself seem important.

[clux]

Exemplar showcase.

Likely to merge this pr over the other for this feature.