First, the reference orbit is in arbitrary precision. Do I just convert it to floating-point numbers in order to use it in the delta iteration?
Yes!
Second, depending on how far you've zoomed in, the delta (i.e. the distance between two pixels) get extremely small, so small that they can't be reasonably represented using floating-point numbers. How does that work?
You can with the double datatype represent values as small as 1e-308.
Can I somehow scale up the deltas in order to work with the larger numbers?
You can multiply the high precision values before converting them to double values with up to 1e308. You then need to divide the values with the same scaling value.
By doing so, you can use ordinary double datatypes up to 1e616, in theory.
In practice, Kalles Fraktaler is using scaling from 1e301 to 1e600 and then the long double datatype is used from 1e600 to 1e4900. And beyond that I use a custom datatype with double as mantissa and integer as exponent. See floatexp in the source of Kalles Fraktaler. (but you could of course use scaled long double to 1e9800)
Third, how do I find out when the iterated value's magnitude is larger than the bailout value? I only have the deltas, so I suppose I have to add it to the reference point and check the magnitude of reference[n]+delta[n] > bailout, but then again, the delta is supposed to be a floating point number whereas the reference orbit is arbitrary precision and not necessarily small, so a lot of precision would be lost when adding the two.
You have already converted the arbitrary precision values to doubles, so reference[n]+delta[n] grows beyond the bailout value. Also the delta value grows from extremely small to a value comparable to the bailout value in magnitude.
And then you will find out that the low precision prevents some areas of a view to be rendered correctly and that glitches will occur in your images. Some of these glitches are one-colored blobs, some with incorrect structures. Then you are ready for the next step, to use Pauldelbrot's glitch detection method to detect these pixels, which need additional high precision references to be rendered correctly.
But you will have a lot of fun!
And don't forget Series Approximation, that allows you to skip a lot of iterations and that is many times even more time saving than what you get from using hardware datatypes instead of high precision.