Optimize ParallelLogic loops slightly

[jtuc]

What

Small optimizations to ParallelLogic.limitByInput/ParallelLogic.limitByOutputMerging

Implementation Details

These changes provide only a small speedup, but the methods are called so often that it was a noticable difference in my testing.
For limitByInput, just tracking the minimum instead of creating a set is an easy way to reduce allocations.
For limitByOutput, we do a similar tweak, but also refactor the loops slightly to avoid double-calling canVoid for each capability. (If canVoid is true for all capabilities, we still return parallelLimit and still avoid calling cap.limitParallel.)

Outcome

Ticks run a tiny bit faster. (It's not a huge difference, so wasn't totally sure it was worth making this PR, but since the code is now also a bit simpler it's probably worth doing.)

Additional Information

I'm not really sure what the intent of RecipeCapability#doMatchInRecipe is - I don't think the logic has changed, but since all the built-in recipes return true I couldn't test this.
I did also look at optimizing the ItemRecipeCapability methods, but I couldn't quite figure out the distinct handling, and the tweaks I did try weren't clear wins in spark.