Should trimming be based mainly on 3′ adapter contamination?

Yes

How do I distinguish between adapter sequence vs low-quality bases?

At the first of sounding flipant, you can tell adapter sequence because it has the sequence of the adaptors, and low-quality bases because the base quality scores are low.

More conceretely, you should trim the reads using a read trimmer. Everyone has their favourite read trimmer. I tend to use cutadapt. Others include trimmomatic, fastx, fastp etc. Give it the sequence of the adaptors (adaptors are fairly standard these days and it should be that hard to find). Most trimmers are configuratble as to the sensitivity of the adaptor matching, and how to set thresholds for low quality bases. I tend to trim any bases below Q25, and matches of 3 bases or more to the adaptors. For miRNA-seq data, I would discard any reads that don't have at least some adaptor sequence.

What is the recommended minimum and maximum length cutoff after trimming for miRNA-seq?

I don't generally use a maximum length, but rather require there to be a 3' adaptor sequence at the end of the read. You could use 30/35 if you like, thats fairly safe, but you will discard miRNA precursors. This may or may not be a good thing. Minimum is also a matter of taste, but something like 15 is probably fine.

Are there any best practices or common pitfalls specific to blood/plasma miRNA-seq data?

The only thing I can think is that plasma is cell free (i think). This means you are going to try to be picking up cell-free miRNAs, and the levels are going to be extrememly low. However, because you will have sequenced many reads, you will get many reads, irrepsective of the number of miRNA molecules those reads represent. This will probably make the data pretty noisey, and, importantly, not comparable really to whole blood or other cell-containing samples.