Ont!

Implementation of `--ont` flag, including disabling of the length correction and 
inclusion of the new ONT, long-read error model.  This PR also includes checks 
for the relevant requirements for long-read alignments (i.e. issues regarding only 
recording certain information in the primary alignment, which doesn't seem to 
be an issue with short read data).

* In the ONT model, the log-likelihood of front clips and back clips are separately calculated based on a semi-geometric model that contribute to the reported log-likelihood.

* Consider alignment with a supplementary alignment as bad alignments
  (probability 0).

* Properly handle very low probability alignments and return the
  sentinel value LOG_0 instead of -inf (as returned by std::log()).

* Use only clips at the front of the alignments for the clipping
  model. Clips at the back of the alignments do not follow a geometric
  distribution.

* Handle hard clip and reordered alignment properly

* Remove debugging code in SalmonQuantifyAlignments.

* Remove unused function recordPrimaryIndex

* Remove logger_ member from ONTAlignmentModel (overshadows the member
  in base class AlignmentCommon)

* Keep track of primary alignment that contains the sequence to use
  with the secondary records that may not contain any sequence
  (happens with long reads to save space).

* Refactor computeErrorCount. Works block by block (instead of base by
  base) for CIGAR operation except 'M' (match, as every base must be
  compared to find mismatches).

* computeErrorCount handles properly hard clip at the last position of
  the CIGAR string.

* Remove some commented code

* Limit some needless indentation

* Simple model based on Binomial distribution and number of errors.

* No distribution regarding soft clip yet.

* Still problems with minimap2 output not including the sequence for
  secondary alignment.

* Simple error model: binomial distribution for errors (where the
  reads are binned by length).

* Two alignment model classes: AlignmentModel (for short read) and
  ONTAlignmentModel (for Oxford Nanopore long reads). Both inherit
  from AlignmentCommon for common aligner task (e.g., handle BAM file
  content).

* Based on the flag --ont, the proper instanciation with one of the
  class above of 'runSingleEndSample' is used. This only applies for
  single-end data without equivalent classes. Single-end with
  equivalent classes and paired-end data do not honor the --ont switch
  (Should there be a warning if the flag is used in that case?)

* refactoring: move large piece of code in the switch base on the
  library type into helper functions (for the single end reads case).

* helper function for single end reads case and no eq classes has
  template parameter for alignment model (right now hard-coded to
  AlignmentModel, the model for short reads).

* All use of AlignmentLibrary either use the AlignmentModel class (for
  short read libraries) or also have a template parameter AlignModelT.