Add explicit RNA reference-completion hypotheses (10.10.0)

[iskandr]

Partial Exacto observations can omit a reference start or stop, so their selected
ORF may describe an internal peptide rather than the protein implied by a
compatible reference continuation. Add the opt-in
reference_completion_hypotheses API to generate those continuations as separate
candidates, preserving the observed RNA, original ORFs and evidence.

Completion works at either mapped end and supports explicitly supplied isoform
alternatives that agree with the observed genomic/exonic path. Each hypothesis
records its assumed spans and observed span, nests original read support, and
translates from the retained annotated start using existing genetic-code and
selenocysteine rules. Known RNA ends prevent extension. Missing starts,
ambiguous coding bases and missing stops remain explicit. No initiation,
translation, expression or extra RNA support is inferred from completion.

For example, observed PGKF remains unchanged while a compatible reference
continuation yields a separate MKPGKFP hypothesis. An unchanged internal-ATG
suffix likewise remains unresolved as an observation; its completed counterpart
can recover the unchanged full reference protein.

Fixes #499, following the observed-coordinate mapping released in #559.
Initiation-dependent upstream extensions remain in #466. Adds user/API docs
and bumps the version to 10.10.0.

Validation:

• 74 new regression cases pass: both strands, either end, split codons,
  in-/out-of-frame producer ORFs, isoforms, frameshifts, RNA ends, ambiguous
  bases, selenocysteine, provenance isolation and real-transcript JSON round-trip.
• Lint, strict documentation build and all 3,070 local tests pass (92% coverage).
  PR and main CI passed on Python 3.9, 3.10 and 3.11; documentation deployment passed.
• Published 10.10.0 from clean main via
  ./deploy.sh; both PyPI artifact hashes match the build. A fresh PyPI install
  passes the CLI smoke check and all 74 new regressions. Tag v10.10.0 is pushed.
• Real Ensembl CFTR/BRCA1 smoke check retains 36 compatible transcript pairs
  from 11 × 29 supplied isoforms.

Scientific inputs checked: pinned Exacto translation,
initiation-context experiments, and
selenocysteine/3′-UTR experiments.