Make matrix algebra (`zero`,`one`,`*`,`+`, ...) work with special matrix types

The correct general behavior for zero and one is to construct a StaticArrays type which can represent the additive and multiplicative identities. Currently this doesn't always work for the reasons described in JuliaLang/LinearAlgebra.jl#1516: for example, zero(Rotations.RotMatrix) should not return a RotMatrix.

One fix would be to make operations always call similar_type and refine the documentation for similar_type to specify that it should return the type of a "general full matrix". "Full" for the purposes of array ops like broadcasting which don't preserve "structural zeros", and full for algebraic purposes because we need a member of GL(n,R) as generic output. This seems consistent(ish?) with similar in Base. Note that changing to use similar_type like this is slightly breaking and led to JuliaLang/julia#700, so this will technically be a breaking change. However we need to do something like this for correctness so we should do it in the next major release. Generally I don't think it will be very breaking...

Perhaps there's some more general option where similar_type can somehow be aware of the algebraic operation... eg, for multiply to call similar_type(m, *, sz, ty)... but I'm not sure that would be any better than simply overriding *.

Begin with the existing implementations of zero, one, *, +, and similar_type, then read JuliaLang/LinearAlgebra.jl#1516 and JuliaLang/julia#700 for the correctness and breaking-change context. Done means the chosen design makes identity and algebra operations work for special matrix types and documents the resulting similar_type contract; no specific file or test is named here.