`toLinear` breaks optimizations

[ulysses4ever]

This has been an issue for a while (previous encounters recorded in #4). Take a look at a profiling build (profiling: True in cabal.project.local) at a12c2df:

cabal run benchrunner -- 1 "Mergesort" Seq 10000 +RTS -pj

(Mergesort is known to suffer more from it but others don't avoid it either). Here's how time profiling looks around get2/size2/setLin with toLinear:

and without it (have to turn the linear arrow -. into the regular arrow -> instead of the linear one -- just fix the definition in Linear/Common.hs):

(the pictures are created by speedscope.app: just upload benchrunner.prof that you get from the above cabal run there).

So, for instance, get2 takes 8 ms vs. 3 ms total time to mergesort a 10K array. This is crazy!