Merge overlapping read pairs into single LocusRead

[iskandr]

If the insert size is small enough then many reads actually overlap the sequence of their mate pairs, which provides an opportunity for merging them into a longer read length. Doing this would also prevent over count of variant allele support.

[iskandr]

May be simplified when we switch to using pysam's fetch instead of pileups (https://github.com/hammerlab/isovar/issues/64) since we can first group all reads at a locus by their template name and then when processing each group merge multiple reads which come from the same fragment.