When using a Bayesian spam filter (like Thunderbird uses), when you mark a particular email as spam it puts every word from the spam into its library of words and marks them slightly more likely to be associated with spam than other words. The more spam you get with certain words, the more likely it is that the spam will be detected as such due to the word frequency.
The point (I've heard) of the gibberish spam, is to clog up your spam filters with words that have no real statistical relation to regular spam, thus rendering your spam filter more and more ineffective and more likely to mark regular emails as spam. Then the real spam is more likely to get through your filter as it's either ineffective or you've turned it off entirely. Thus, it may be wise to not mark gibberish-spam as Junk Mail, as it won't do you any good and may potentially render your filter less effective over time. When they send out the gibberish-spam, they're thinking long-term benefits.
