Implement 1s pause between mutations

[ericphanson]

closes #232

Based on #233 which is why the diff is so huge here. That one should be merged first

[KristofferC]

The LAST_MUTATION_TIMESTAMP should per per API token or something like that, or? It feels somehow the wrong abstraction level to make this a single global of the whole package.

[ericphanson]

True, it should be at the same granularity at which the rate limits are applied, which is per account iiuc. Per token seems like a reasonable approximation. I’m not sure a good way to do that, it would feel weird / insecure to make a global dictionary with tokens as keys. We could hash them or something. Or use the auth object itself (add a mutable field), but then if you created 2 auth objects with the same token they wouldn’t share timestamps.

[ericphanson]

ChatGPT suggested the key to be

reinterpret(UInt64, sha256(token)[1:8])

and cleaning up old entries to not accumulate memory forever in a long running process with many tokens used.

[ericphanson]

ok, it adds a fair bit of complexity to track the state across auth instances but I think the current state does so:

• NEXT_AVAILABLE::Dict{UInt64, Float64} maps token-hashes to availabilities (i.e. next time that auth is allowed to make a mutation)
• cleans up old key-value pairs periodically to prevent memory leaks
• aims to be not-terrible in concurrent scenarios: global data structures protected by locks, uses reservations instead of last-used-timestamps so tasks wait the correct amount of time instead of looping 1s waits until ready

[DilumAluthge]

Alternate pitch: Instead of tracking the timestamp of the last mutating call, just always sleep(Second(1)) before every mutating call? It would simplify the implementation significantly, and is probably fine for most users?