Make sam_hdr_remove_lines rebuild the target arrays.

[jkbonfield]

Without this, attempting to remove an SQ line works for SAM and CRAM, but BAM keeps it in the binary portion of the header and the SQ line is put back again on decode. It'd likely have also broken sam_hdr_tid2name.

[jkbonfield]

Switched to draft. The PR may be trivial and it may well pass tests, but it's breaking samtools.

It turns out that removing SQ lines does need this if it's to fix the headers written out in BAM, however in BAM world it then can no longer decode some records (SAM works) via the multi-region iterator. I suspect it's still doing the name lookup for RNEXT and failing (if we remove a ref referred to in another record). In SAM this produces a warning, and in BAM it's a hard failure. We need to be consistent in error handling.

[jkbonfield]

I was overly cautious earlier on seeing my samtools failures.

With hindsight it's obvious. If we modify a header then the in-memory binary bam1_t format is invalid as we read it from the input (with full headers) and we need to correct tid and mtid before writing it back out with the shrunken header. This fix is therefore fine and the problem lies in my own samtools changes.

We do still have inconsistencies in how we handle invalid references, but that's baked into the format. In SAM it's text so we can validate. In BAM it's binary so while we could check ranges in more places, we can't spot most invalidities. It's also an issue for another PR should we wish to fix it.