Filtering out features with low overall counts sounds like a good idea to reduce memory requirements for the job.
So, you can filter features with counts less than a certain threshold (50,100 or other numbers), filter sequences based on filtered feature table and then align sequences. In my experience it will decrease memory requirements and speed up the process.
Not necessarily. For core-metrics, one need to rarefy the data, but it is already implemented in the plugin. You don't need to rarefy for other plugins unless it is specified in the plugin options.
I would do it in that order.
Usually for bacteria I filter out all features with absolute counts less than 50, but you also can filter based on relative abundances (like less than 1%, 0.1%) and prevalence.