`InjectionInfo.parse` can be very expensive and in the case of some MixinExtras features is not pure.
The correct approach would be to use `InjectionInfo.getInjectorAnnotation`, but the override is no longer needed anyway since we only support Java 8+ so Mixin will already make injector methods private and synthetic.
Except, Mixin does it based on the current `COMPATIBILITY_LEVEL`, which is wrong, because we could be on JAVA_17 and yet a class compiled with JAVA_8 had no choice but to use public methods.
The original implementation was also incomplete because it forgot about default methods.
The old one did not properly account for discontinuities in the list, leading to double-width types clobbering meaningful locals and causing mysterious VerifyErrors.
These will never match because the LHS is the mixin name and the RHS is the target name.
This restores the 0.8.5 behaviour which does not check the owner at all.
We have never supported assignment expressions in initialisers until now anyway, and even if we check the owner we cannot determine whether the instruction is operating on the current *instance* without proper static analysis, so IMO there's not much point bothering.
Bafflingly, Mixin fails to match a `NEW` instruction whose `<init>` call is outside the slice range, *iff the user specified a descriptor*.
I'm even more surprised that a mod relied on this.
Fixes#152.
The old logic doesn't really make any sense and the calculated extension usually ended up being negative.
Instead, we just make room for our possible receiver and all of our args.
The only reason MixinExtras uses this method is to match Redirect's behaviour, so I would like it to be kept in line with that.
Additionally, there is nothing wrong with not checking the desc as long as you've definitely found the right `NEW` insn, which the injection point handles itself.