Since you have mixed-orientation reads, I would recommend using the vsearch or blast-based classifiers in q2-feature-classifier, which can handle mixed-orientation reads (i.e., these perform both forward and reverse alignment against your reference sequences). We do not yet have a :qiime2: method for re-orienting reads, so just using an appropriate classifier would be the easiest thing to do — and those classifiers do perform quite well also.
There may be ways to perform read re-orientation outside of QIIME2, e.g., see this post for another user's reorientation script. However, I would personally not go that route, if only to avoid going outside of QIIME2 (to preserve provenance, avoid formatting snafus, etc).
Let me know if you have any more questions or concerns!