Basically, the strength of passwords can look really good on paper (96^12 possibilities or whatever), but in practice the entropy is much less: individual characters are not as independent as we'd like to think, and your clever word or phrase may be in a dictionary already (or may be a short Levenstein distance from something in a dictionary).
If guys using vanilla hardware get that kind of success in 1 hour with MD5, you only need to increase hardware and the time required to see it's still completely doable for other hash functions.
If for some reason you can't bring the bcrypt code into your project, you can implement pbkdf2 using basically a loop and your stdlib's hash functions.
Scrypt takes the whole concept of placing extra demands on the computer and applies it to the RAM rather than the CPU (perhaps as well as?,) the idea being that RAM is harder to accumulate in obscene quantities than CPU power.
The algorithm uses CPU cycles to generate the hash. So there is no way to speed it up, just by using a faster computer.
An even better option is scrypt, which is memory based, which is much harder to scale.
Using MD5 hashes was done as a relaxing constraint -- at the other end, the attackers had tightened constraints by getting limited time and computing power.
The goal was to show how attackers work and think.
> The list contained 16,449 passwords converted into hashes using the MD5 cryptographic hash function.
Edit: I'm removing all my snarky nitpicking. This is a good article. Yes, the MD5 case they present is a poor case, but it's really about demonstrating the tactics of attack selection, rather than teaching someone how to make crack-resistant password schemes.