Revamp Mul, Ldiv, and Inv as Applied

[dlfivefifty]

A more general approach is to support general application of functions. The current lazy operations could be type aliases.

[dlfivefifty]

This also takes care of matrix functions like lazy exp(A)

[dlfivefifty]

@tkf This would mean we have the option of the following type aliases:

const Mul = Applied{typeof(*)}
const Ldiv = Applied{typeof(\)}
const PInv = Applied{typeof(pinv)}
const Inv = Applied{typeof(inv)}

const Plus = Applied{typeof(+)}

The last one is tricky because we have the choice between Applied and Broadcasted because they mean the same thing here. I forgot why we decided to leave Plus as a BroadcastArray{typeof(+)} instead of a Broadcasted, but maybe Applied solves the same problem?

[tkf]

Cool! It looks like LazyArrays is becoming a fully customizable computation graph library.

Re: Plus. I'm not sure what's the best way, but here is an observation: ones(3,3) + I works but ones(3,3) .+ I does not. Lazy expression involving UniformScaling sounds very useful. Doing so with Broadcasted would be incompatible with standard broadcasting implementation (at least ATM). So, maybe Applied is better so that we can do more advanced stuff without breaking the semantics of Broadcasted?