How to find it: do MD5 on some random input. Then do MD5 on that output, and repeat until it converges. The problem is that you could end up in a cycle, instead of at a fixed point.
The speed with which it can be found: I'm going to wave my hands and claim this is in "Analytic Combinatorics" by Flajolet and Sedgewick. Seriously, though, under the assumptions we've been throwing around here this is a "random mapping" and these are reasonably well-studied objects.
It would be interesting to try to map the group properties of MD5-space. Every 128-bit number would either be in a slide to a cycle or in a cycle (a fixed point is a cycle of size 1).
This still doesn't change the huge-normousness of the task, though.
No need to worry about cycles from using one output as the next input; just iterate over all possible 128-bit inputs, in any order. Unless you're trying to leverage some known deviation of MD5 from being a true random oracle, each input is just as likely as any other to be an identity input.
The speed with which it can be found: I'm going to wave my hands and claim this is in "Analytic Combinatorics" by Flajolet and Sedgewick. Seriously, though, under the assumptions we've been throwing around here this is a "random mapping" and these are reasonably well-studied objects.