Bungee doesn't resend a join game for old servers, so we will have to keep data stored there across server switches
This likely fixes other issues with chunk data writing after server switches on legacy servers as well
Defaulting to submitting to the netty event loop caused issues more often than not - this also removes the `currentThread` flag and instead provides new scheduleSend methods so it is always obvious whether the packet is sent immediately.
This essentially merges the two approaches to the metadata handling from ViaVersion and ViaBackwards and improves on both designs.
ViaVersion did not track every single entity, but only those needed (at least in theory) and can work with untracked entities' metadata. It had a very simple method overridden by metadata rewriter implementations, directly operating on the full metadata list and manually handling meta index changes as well as item/block/particle id changes.
ViaBackwards on the other hand had to track *every single* entity and threw warnings otherwise - while less prone to errors due to giving obvious warnings in the console, it unnecessarily tracks a lot of entities, and those warnings also annoys users when encountering virtual entity plugins (operating asynchronously and sending update packets while already untracked or not yet tracked). Dedicated MetaHandlers made id changes and filtering a lot easier to read and write. However, the actual metadata list handling and its distribution to handlers was not very well implemented and required a lot of list copying and creation as well as exception throws to cancel individual metadata entries.
This version has MetaFilters built with a Builder containing multiple helper functions, and the entity tracking is properly given its own map, hashed by a Protocol's class, to be easily and generically accessible from anywhere with only a Protocol class from the UserConnection, along with more optimized metadata list iteration. The entity tracking is largely unchanged, keeping ViaVersion's approach to not having to track *all* entities (and being able to handle null types in meta handlers).
All of this is by no means absolutely perfect, but is much less prone to errors than both previous systems and takes a lot less effort to actually write. A last possible change would be to use a primitive int to object map that is built to be concurrency save for the EntityTracker, tho that would have to be chosen carefully.
This has been tested on the following:
- AdoptOpenJDK Java 1.8.0_282
- GraalVM CE 21.0.0 OpenJDK 11.0.10
- AdoptOpenJDK Java 15.0.2
- AdoptOpenJDK Java 16 (also tested with BungeeCord b1556)
- Amazon Corretto OpenJDK 16.0.0.36.1
... with Waterfall b406 on Linux 5.10.28.
