Rename NonEmpty.groupWith/sortWith functions to *On*

Motivation

Referring to these functions: sortWith groupWith groupAllWith groupWith1 groupAllWith1.

The On suffix is more common for such things, and I anticipate it's what most users will normally expect them to be named.

Additionally, the name groupWith is potentially misleading. That name is also used in GHC.Exts, for a function with the same type signature but different semantics. GHC.Exts.groupWith sorts its input list, unlike NonEmpty.groupWith or Data.List.Extra.groupOn. (In NonEmpty, we have groupAllWith for group-with-sort. So we have three names, two semantics, and no consistency.)

According to https://github.com/ekmett/semigroups/pull/52, With was chosen because:

The preferred vocabulary for On here is With to match the combinators in GHC.Exts from the "comprehensive comprehensions" paper. That'd make it so that if you use these with comprehensive comprehensions turned on and RebindableSyntax turned on then you'd get these combinators.

But I don't see anything in the docs which suggests that the TransformListComp extension uses these functions implicitly, or interacts with RebindableSyntax. So my guess is this is no longer a concern.

The case for sortWith is weaker, since it might trip people up but it won't introduce unexpected semantics. There's also a counter argument in #12044 (closed). Data.List.sortOn only computes the mapping function once for each element, while GHC.Exts.sortWith computes it every time it does a comparison but uses fewer allocations. Data.NonEmpty.sortWith acts like the latter. My suggestion would be to replace it with a sortOn implemented like Data.List.sortOn, but I also don't think it would be terrible to have both. If there's no agreement on this function, I think it would be worth renaming the group functions anyway.

Proposal

Rename these functions to have "On" instead of "With". In the case of sortWith, also reimplement it as

sortOn :: Ord b => (a -> b) -> NonEmpty a -> NonEmpty a
sortOn = fmap snd . sortBy (comparing fst) . fmap (\x -> let y = f x in y `seq` (y, x))

I assume the process here would be to leave the existing names for a time with deprecation warnings, and then remove them a few versions down the line.