fix(NODE-3769): retryable writes are not compliant with specification

[nbbeeken]

Description

Fix NODE-3769, test NODE-2111

What is changing?

• Add a new MongoUnexpectedServerResponseError which is a subclass of MongoRuntimeError I could continue using MongoRuntimeError to avoid adding a new class, but the refactor/fix to execute operation called for it.
• Refactoring / correcting of executeOperation's retry logic
• isRetryableWrite now only checks the label if we're connected to a topology above a certain wire version, as required by the spec.
• Refactored the retryable write runner, no logical changes, just async/await, I ran my changes against the unmodified runner on main, so I'm confident that my test runner changes are safe
• Moved the labels to an enum MONGODB_ERROR_LABELS, This allows us to type check our labels, (mostly could catch future spelling mistakes if we miss actual test coverage) felt like it was still relevant to this work to keep in.
• Deleted the unused clearAndRemoveTimerFrom function. entirely internal, never reached.
• Deleted connectServers called only from one place and simply created a level of indirection that is frustrating to debug
• Deleted isRetryableEndTransactionError, it had one condition, checking for the retryWriteLabel, inlining the condition has better readability, there could be reason to bring it back if there is more conditions.
• Expanded the conditionals in supportsRetryableWrites and made it accept an optional server (removes an existence gate from a caller if stmt)
• Removed the constrained timeout from LB testing, causes a lot of timeout failures when you debug locally
• async-ify testOperations in the legacy spec runner, debugging.
• Added to our error unit testing, we had a public error that wasn't exported at the top level, I think my test should catch all those cases now.
• New instructions for launching LB cluster
• ran prettier on global.d.ts, it doesn't run automagically currently 😿

What is the motivation for this change?

Our retryable logic was falling back on error codes, the spec explicitly states that on newer server versions the label should be the ONLY source of truth. So we were still retrying errors, and we will continue to do so, but the rules were incorrect.

Double check the following

• Ran npm run check:lint script
• Self-review completed using the steps outlined here
• PR title follows the correct format: <type>(NODE-xxxx)<!>: <description>
• Changes are covered by tests
• New TODOs have a related JIRA ticket
  • TODO can we break the type of hasErrorLabel, addErrorLabel, narrowing from string to known labels
  • TODO file ticket about try catch in execOperation
  • TODO inside commandHandler of endTransaction is it okay that we only check the retry read condition before unpinning / adding the unknown commit label? I don't think so, but also the retry has already occurred at this stage, so I don't think the session gets reused, unpinning might be correct.

[dariakp]

Partial feedback so that I don't lose it, still got the main files left to go through

[dariakp]

Some partial feedback on the revisions

[dariakp]

Just the one additional comment here

[dariakp]

Putting an approve because these are minor: