When passed a multidimensional array, simply iterate over its columns, independently filtering each one.  The starting filter state may be given as a vector (in which case it is common across all columns) or an array with the same number of columns as the input.

Previously, filt[!] methods were only defined for arguments of all the same type T<:Number. This relaxes that restriction, allowing one to, for example, filter an array of Float32s by Float64 coefficients.

By default, the output of filt will be the determined by promote_type of all arguments.

Instead of defining many type parameters (T<:Number, S<:Number, ...), filt now simply trusts that the eltype of the arrays have methods for basic arithmetic.

Because we're back to only one optional element for both `filt` and `filt!` (the initial state), I figure it's simpler to just use a positional argument in both cases.  So now the API is:

    filt(b, a, x, [si])
    filt!(out, b, a, x, [si])