`norm` of nested `StaticArray`s throws

[thchr]

The implementation of norm(a::StaticArray) assumes that the elements of a have an abs2 implementation, i.e. that they are scalars. Unfortunately, that means that if a's elements are also e.g. SVector, it will throw:

julia> a = SVector{2}(SVector{2}(0.0, 0.0), SVector{2}(1.0, 1.0))
2-element SVector{2, SVector{2, Float64}} with indices SOneTo(2):
 [0.0, 0.0]
 [1.0, 1.0]

julia> norm(a)
ERROR: MethodError: no method matching abs2(::SVector{2, Float64})
Closest candidates are:
  abs2(::Bool) at bool.jl:80
  abs2(::Real) at number.jl:151
  abs2(::Complex) at complex.jl:265
  ...
Stacktrace:
 [1] macro expansion
   @ ~/.julia/packages/StaticArrays/rKBnO/src/linalg.jl:225 [inlined]
 [2] _norm
   @ ~/.julia/packages/StaticArrays/rKBnO/src/linalg.jl:213 [inlined]
 [3] norm(a::SVector{2, SVector{2, Float64}})
   @ StaticArrays ~/.julia/packages/StaticArrays/rKBnO/src/linalg.jl:212
  ...

One of the unhappy consequences of this is that things like isapprox breaks for nested StaticArrays.

The norm implementation is here:

StaticArrays.jl/src/linalg.jl

Lines 213 to 228 in d03a5e6

	@inline norm(a::StaticArray) = _norm(Size(a), a)
	@generated function _norm(::Size{S}, a::StaticArray) where {S}
	if prod(S) == 0
	return :(zero(real(eltype(a))))
	end
	expr = :(abs2(a[1]))
	for j = 2:prod(S)
	expr = :($expr + abs2(a[$j]))
	end
	return quote
	$(Expr(:meta, :inline))
	@inbounds return sqrt($expr)
	end
	end

AFAICT LinearAlgebra is avoiding this by calling out to an internal function norm_sqr instead of abs2, which is just abs2 for scalar input, but recursively calls norm otherwise. There's actually a norm_sqr function in StaticArrays too, but it seems unused?

If the suggested fix seems acceptable, I can make a PR.