Here we go. This PR looks massive (and was a lot of work), but a great deal is just test cleanup. The PR is based on #4684 and #4678, so look only at the last 2 commits when reviewing.

• Connection-level type mapping is now obsolete-as-error (we point users towards the release notes, which will detail data source mapping).
• I also obsoleted global type mapping (not as-error), since they're inferior and we want to nudge people towards data source (but see #4690). This may be a bit premature/extreme though (e.g. replication doesn't yet support data source).
• Note that global mapping changes now only affect connections before their data source is created. Previously global mapping changes took effect when a connection was closed/open.
  • This is a breaking change. Though most people were just putting global mappings at the start of the program (this is fine), if a user globally mapped late - after their pool was already created - this will now break.
• Since the DatabaseInfo is now held/cached by the built NpgsqlDataSource, I've removed the global cache for it. That means that if you build two data sources to the same database, we do type loading twice. Previously the caching was necessary since each individual connection open would load it otherwise, but with data sources I don't think it matters. But I did it in a separate commit in case you guys think we should keep the caching.
• We may want to add ReloadTypes/ReloadTypeAsync to NpgsqlDataSource (since it works at the data source level). If so, we may also want to obsolete it at the connection level, though there's some utility in the specific connection also being refreshed immediately.
• Naming-wise, things are now a bit weird: the (internal) TypeMapper does not, in fact, implement INpgsqlTypeMapper; that interface is about letting users manage mappings/plugins, and is now implemented by GlobalTypeMapper and NpgsqlDataSourceBuilder only. But it doesn't seem important.

Closes #4494
Does part of #4572