It is continually surprising to me that there are all these reverse engineered minecraft servers yet none of them seem to provide a significant performance advantage within a single world. Cuberite, Glowstone, etc. They provide a notable boost when your machine is serving many worlds, but it seems most still use a single thread per world model. At least cuberite offloads chunk generation to a second thread, but in their own documentation they claim they don't want to support arbitrary levels of multithreading for fear it would hurt performance on anything less than a quad core. Even though this is a bit of a misconception, I don't understand why they wouldn't make such a thing configurable.
As an aside they make a claim that will be familiar to many lisp programmers. Their claim is by providing better power they will improve the average performance in the long run for their users (e.g. people who like lots of mods). Look at it this way, if you use horribly buggy edge cases to build flying machines that's going to take a ton of work for the server to simulate (minecraft flying machines often use tick perfect timings arrived at through - effectively - busy loops); if instead you used a mod that provided flying mechanics, and flying control blocks (e.g. that knew when to inject their logic) you would save a lot of performance. They may not be able to have the fastest server around, but given the things their users will use it for, theirs will probably be faster than those efficient servers in those cases.
Even if you conceive an API from scratch, Java hasn't the tools to prevent plugin authors from doing things they shouldn't in this regard.