Usually, in Casava >=1.8 format, I look for consecutive set of bases, for each read, with quality "#". This is similar to the "B" in older formats. For example:

@HWI-ST863:108:C0F9BACXX:8:1101:1235:1979
CTTACTAAAGACAATGGTGTTCTTCTCATCTTCGATGAAGTCATGACTGGATTTCGTCTAGCCTATGGTGGAGCCCAAGAATACTTTGGAATCACGC
+1:N:0:ATCACG
CCCFFFEFHHHHHJJJJFEIIJJJGHIJJHIJJJIJJIEHIIIJIIJIJJGIIIJHIGH######################################

I trim to obtain:

@HWI-ST863:108:C0F9BACXX:8:1101:1235:1979
CTTACTAAAGACAATGGTGTTCTTCTCATCTTCGATGAAGTCATGACTGGATTTCGTCT
+1:N:0:ATCACG
CCCFFFEFHHHHHJJJJFEIIJJJGHIJJHIJJJIJJIEHIIIJIIJIJJGIIIJHIGH

This has resulted in an average quality >= 30 for all my fastq files. When I mean average, its not per read, rather, the average quality at each base over all reads. If the trimmed sequence length is > a threshold ( 60 for 100 base reads in my case), then I keep them. Otherwise, I discard them. Just to give an idea. This can be done separately for both pairs. If one of the pairs gets discarded while trimming for quality, then I keep the other as a single end read in a separate file. After quality trimming, I balance the pairs so that each pair is on the same line between the 2 fastq files. However, I don't understand how/why the paired end reads are of different lengths.

Trimming based on length is probably going to be less than optimal since you will likely end up cutting off some high-quality bases in some reads and leaving low-quality bases in others. I'd suggest using any one of a dozen or more trimming softwares (cutadapt, fastx, sickle, trimmomatic) to do the quality trimming based on individual read base qualities.