Add rrule for infinite-norm

[sethaxen]

The current rrule for norm fails when p=Inf, for which the norm is equivalent to maximum(abs, x):

julia> using LinearAlgebra

julia> x = Float64[1, 2, -3];

julia> norm(x, Inf)
3.0

julia> _, back = ChainRules.rrule(norm, x, Inf);

julia> ∂self, ∂x, ∂p = back(1.0);

julia> ∂x() # expected: [0.0, 0.0, -1.0]
3-element Array{Float64,1}:
   0.0
 NaN
 NaN

julia> ∂p() # expected: 0.0 or Zero()
NaN

This really just fails for any very large p (e.g. also fails for p=1e3). It works fine for p=100:

julia> norm(x, 1e2)
3.0

julia> _, back = ChainRules.rrule(norm, x, 1e2);

julia> ∂self, ∂x, ∂p = back(1.0);

julia> ∂x()
3-element Array{Float64,1}:
  5.820975652447899e-48
  3.689481639869744e-18
 -1.0

julia> ∂p()
0.0

The actual implementation of norm in LinearAlgebra special-cases p of 0, 1, 2, -Inf, and Inf, falling back to normp if another p is provided (see https://github.com/JuliaLang/julia/blob/master/stdlib/LinearAlgebra/src/generic.jl#L604-L619). Should the rrule for norm here instead be replaced by rrules for these functions? I see 2 downsides: 1) more code 2) pullback of p for these special p's would be Zero(), and in general would not be type-stable when pulling back through norm.
On the upside, the pullbacks for these types can be implemented a little more efficiently.