fix: four bugs in integrate and exon_fasta_to_twobit modules

[anijudy]

Found and fixed four bugs while running TOGA2 on RHEL8 (glibc 2.28)
with three references against a query genome.

exon_fasta_to_twobit.py

• --fatotwobit_binary silently ignored: Line 166 hardcodes
  f"faToTwoBit {tmp_path} {output}" instead of using self.fa2twobit,
  so the --fatotwobit_binary CLI flag has no effect at the SLEASY step.
  Fixed to f"{self.fa2twobit} {tmp_path} {output}".

integrate.py (prepare_sequence_files / _read_fasta)

• gzip read mode: gzip.open was called with "rb" (bytes), but
  _read_fasta uses str.startswith, causing
  TypeError: startswith first arg must be bytes or a tuple of bytes, not str.
  Fixed to "rt".
• projection lookup missing reference prefix: self.final_projections
  stores names as "{species}.{projection}" (e.g. hg38.XM_011510358.3#NBAS#328),
  but _read_fasta looked up the bare projection name without the prefix —
  so nothing ever matched and protein.fa.gz / nucleotide.fa.gz were
  always empty. Fixed to check f"{ref}.{proj}".
• gzip write mode: the write step used open() instead of gzip.open(),
  producing plain-text files with a .fa.gz extension (invalid gzip).
  Fixed to gzip.open(..., "wt"/"at").

constants.py (Nextflow template)

• DSL2 incompatibility: the joblist guard and Channel.fromPath call
  were at the top level, which is not valid in Nextflow DSL2. Moved both
  inside the workflow {} block.

[ReverendCasy]

Thank you very much for your contribution. Sequence extraction for the “integrate” mode has been mostly tested with the BigBed input, so reading from FA did not get enough attention so far.