Client store generalization

[Joerger]

This PR generalizes client store logic to enable more operability between client backends (~/.tsh, identity file, in-memory, agent forwarded keys). This lays the groundwork for #19421.

• Replace the old KeyStore with ClientStore, which is made up of a key store, trusted certs store, and profile store.
  • Add new KeyStore interface with an FS and in-memory implementation, based on original key store.
  • Add new TrustedCertsStore interface with an FS and in-memory implementation, based on original key store.
  • Add new ProfileStore interface with an FS and in-memory implementation.
  • Each individual store can have a different backend (FS or in-memory). This is useful, for example, with tsh --add-keys-to-agent=only, which keeps keys in memory while writing trusted certs and profiles to disk.
  • Identity files are now used to populate a full in-memory ClientStore on tsh start.
  • Remote agent forwarding sessions will also use a full in-memory ClientStore populated from forwarded keys.
• Fix inconsistencies in Trusted Host certificate handling across different types of keys and key stores.
  • auth.TrustedCerts.ClusterName should never be empty so that it can actually be useful for logic.
  • auth.TrustedCerts.AuthorizedKeys, previously inaptly named HostCertificates, should always be in authorized_keys format rather than known_hosts format to maintain consistency across backends. We only convert to known_hosts when writing to ~/.tsh or identity file.

[marcoandredinis]

Partial pass
Looking good 👍

[marcoandredinis]

Most changes look good 👍
There are a lot of changes that are just moving around methods/structs to different files

I wonder if grouping those changes together would make this PR easier to review

[jakule]

nit: I'm not sure about the performance differences, but you could use DeduplicateAny() in the implementation here:

func Deduplicate[T comparable](in []T) []T {
	return DeduplicateAny(in, func(t T, t2 T) bool {
		return t == t2
	})
}

[Joerger]

Deduplicate[T comparable] has better performance since it can use a map[T]bool for O(1) existence checks, so I think it's better not to generalize all uses to the current DeduplicateAny with O(n) existence checks. Let me know if you disagree since the performance is pretty negligible to be honest.

I was considering changing DeduplicateAny to func[T any](in []T, hashFunc func(T) string) []T so Deduplicate could simply call that, but I don't think it's worth the extra complexity right now.

[jakule]

nit: I'd be nice to also unit test the Deduplicate()

[Joerger]

These tests already exist above here, no change was necessary for the generification.

[ravicious]

It looks like this change broke adding in a new cluster in Connect. clusters.Storage.addCluster calls client.Config.SaveProfile directly on client.Config instead of client.TeleportClient, without setting ClientStore first.

I fixed that in #20263 and added an integration test which would've caught this issue.