Disable k-mer heuristic in some cases

Start with the discussion carried over from issue #663 and review how the k-mer heuristic affects anchored, file-based adapters such as -g ^file:adapters1.fa -G ^file:adapters2.fa. Compare the proposed opt-out and selective-application approaches, then benchmark the amplicon-like command pattern to establish what behavior and speed improvement would count as done.

Hi - I was excited to see this upgrade as we use cutadapt extensively - it works great and can do things no other software of its kind can do, but is one of the slower components of our pipeline, so we're always interested in speed improvements.

However as it turns out we actually observed a mild decrease in speed (~10%) relative to earlier versions. Our use case is mainly trimming 5' adapters from "amplicon-like" data, whereby we load files of thousands of anchored adapter sequences (-g ^file:adapters1.fa -G ^file:adapters2.fa), and every read is expected to find a match to at least one sequence in the file (and possibly non-anchored matches to several). I'm not a software engineer but I wonder if for this use case the cost overhead of the new heuristic isn't worth it (even though most adapters won't be present in any particular read).

Would it be possible to either make this functionality optional via a switch, or else automatically turn it on selectively depending on the alignment mode and/or adapter characteristics? E.g. maybe it makes more sense for for non-anchored adapters vs. anchored (since adapters are more likely to be present in applications that involve anchored adapters, like primer sequences). Or perhaps when the adapters are too short relative to the error rate, the rate of coincidental matches in reads is too high; for a 20 nt adapter sequence with an error tolerance of 10% (2 errors), the kmers are so short that coincidental matches may not be rare.

The speed decrease isn't huge, and the new software version includes other useful features; it's just a bit of a bummer that at least for us the new heuristic is a cost not a benefit.

Originally posted by @eboyden in https://github.com/marcelm/cutadapt/issues/663#issuecomment-1483998923