single-chunk and multi-chunk runs with streaming arbitrary resampling don't match

[ssfrr]

There's something weird going on with the streaming resampling FIR filter with a floating-point ratio. I'd expect to get the same result (modulo float precision) whether I resampled an array all at once or in chunks with the stateful FIR filter. In the below example it looks like that's not happening. It doesn't happen if I change the ratio to a rational number, so it seems to be a problem just with the FIRArbitrary kernel

function runtest()
    ratio = 9000/48000
    data = rand(64)
    f = FIRFilter(resample_filter(ratio), ratio)
    out1 = filt(f, data)
    f = FIRFilter(resample_filter(ratio), ratio)
    out2 = vcat(filt(f, data[1:32]), filt(f, data[33:end]))

    @assert length(out1) == length(out2)
    for i in 1:length(out1)
        if !isapprox(out1[i], out2[i])
            println("mismatch at element $i: $(out1[i]) != $(out2[i])")
       end
    end
end

julia> runtest()
mismatch at element 7: 0.0021295247679830086 != 1.04360682293e-312
mismatch at element 13: 2.2729773433e-314 != 0.013122266681143116

julia> runtest()
mismatch at element 7: 0.002004266351359534 != 3.254494040945382e-86
mismatch at element 13: 2.2729773433e-314 != 0.012160209855406971

julia> runtest()
mismatch at element 7: 0.001037523637825713 != 1.4004377841750797e195
mismatch at element 13: 2.2729773433e-314 != 0.0076082268936907005

There's a comment in stream_filt.jl above function inputlength(kernel::FIRArbitrary, outputlength::Integer) that says

TODO: figure out why this fails. Might be fine, but the filter operation might not being stepping through the phases correcty.

So maybe that's a clue...

[wheeheee]

Closing in favour of #317; runtest now produces a BoundsError as well.