[proxy] Migrate to async

[funbringer]

This makes most of the proxy async (except for mgmt, but we expect to drop it soon). The auth logic is left almost untouched, since this patch is big enough, and it would be good to merge it and unlock further changes.

[funbringer]

I've just realized that I forgot about NUM_BYTES_PROXIED_COUNTER. While it's quite easy to restore this logic, I'm a bit hesitant, since the current approach means that all runtime threads will compete for atomic var's cacheline. I might be overthinking, because there's an awful lot more happening inside tokio's runtime. Thoughts?

[bojanserafimov]

I've just realized that I forgot about NUM_BYTES_PROXIED_COUNTER. While it's quite easy to restore this logic, I'm a bit hesitant, since the current approach means that all runtime threads will compete for atomic var's cacheline. I might be overthinking, because there's an awful lot more happening inside tokio's runtime. Thoughts?

There is some cache contention but we only update when we flush, so I wouldn't expect visible performance impact. If there is any, we can just buffer the metric updates even more aggressively.

[LizardWizzard]

Nice! Great Work! Having this covered by unittests is a huge plus.

What do you think about hypothetical Proxy struct that can hold reqwest::Client for console interaction and a cancellation keys map? This way we can also avoid global variables. May also simplify some future tests that manipulate this state.

I've left just a few minor nitpicks

[funbringer]

@LizardWizzard

What do you think about hypothetical Proxy struct that can hold reqwest::Client for console interaction and a cancellation keys map?

Could you elaborate? I don't see why they should live in a single map.

This way we can also avoid global variables.

How so? You never said anything about this map's lifetime. Besides, I've just re-introduced some private global maps exactly for the sake of simplicity.

May also simplify some future tests that manipulate this state.

Not necessarily.