fix type stability in all norm variants

[stevengj]

Continuing where JuliaLang/LinearAlgebra.jl#91 and #6057 left off, this makes all the norm functions type stable, both internally and externally.

Previously, similar to JuliaLang/LinearAlgebra.jl#91, norm(x, p) was not type stable, in the sense that the norm of an integer matrix or scalar would return different types for different p. It was also not internally type stable, in that the reduction loop would change the type of the accumulator variable if typeof(zero(T)) != typeof(zero(T)+zero(T)), similar to #6069.

Now, all matrix norms return a floating-point result, and accumulations are performed in at least double precision, similar to vecnorm. norm(x::Number, p) returns a value with the same type as real(x) or abs(x), since adding the floating-point conversion there would have complicated the code for no useful purpose.

I also made one or two other code cleanups. The p parameter is consistently declared to be Real, and the dense vector norms override vecnorm1 and vecnorm2 rather than vecnorm to exploit dispatch.