Avoiding idle threads during window encoding in dorado polish.

[lorenzwidmer]

Fixes #1503.

[lorenzwidmer]

There is another bug that can cause low GPU utilisation and has been fixed in the last commit (a6d38c4). I think this is the underlying cause for us really noticing the slow encoding in some cases. In the worst case, it can prevent all threads from encoding for several minutes if the reference file contains regions that are not found in the alignment.

The problem is that bam_itr_querys() can return a null pointer rather than an iterator if the region is not found in the BAM. This is not checked in dorado::secondary::calculate_pileup() and can result in high IO usage and very long pileup durations for non-existent regions in large BAM files (or BAMs with a large header).