Use proquints for recovery codes

[espadolini]

The github.com/sethvargo/go-diceware library that we use to generate friendly recovery codes unconditionally causes a ~1.5MB allocation at runtime, since it uses a map in a var; while this is a relatively small cost for a long running teleport process, this cost is incurred even by binaries that are intended to be much leaner (such as tsh or tbot) and that might end up running in multiple copies (hundreds or thousands) on a given machine. That fixed cost also happens on every new re-execution of teleport to handle shell execution, port forwarding and various other bits of functionality that must happen in a separate process - which is pretty silly, considering that most of those are not in the auth service, which is the only component that needs to generate recovery codes.

Adding to the aforementioned issue, #5685 is still very much a problem, and trimming one more dependency is worth in and of itself.

This PR thus proposes to change the recovery code generation to use proquints (PRO-nouncable QUINT-uplets of alternating unambiguous consonants and vowels, arXiv:0901.4016 [cs.SE]), which we've been successfully using in Cloud to generate random pronounceable identifiers for a while now. As a bonus, this makes the recovery codes slightly less guessable (from ~103 to 128 bits of entropy each).

Example of the account recovery UI after the change:

Existing recovery codes are stored hashed in the user recoverycodes resource, and will continue working as normal after this change.

[zmb3]

🎉

[public-teleport-github-review-bot]

@espadolini See the table below for backport results.

Branch	Result
branch/v13	Failed
branch/v14	Failed
branch/v15	Failed