Make adaptive transform a Vcat instead of a cached vector?

[dlfivefifty]

Sometimes we use the transform to build operators, and the bandwidth is determined by the number of non-zeros. In this case, cached vectors are not so useful as anytime we access them they grow, and the bandwidth of the operators grows.

But is there any reason we would need coefficient vectors to be mutable (in their zeros) by default? A user could call Base.copymutable(P \ f) if they really wanted to mutate.