refactor: uniformize somatic-variant definition across bin scripts

[m-huertasp]

Small PR.

What this does

Determines single definition of a somatic variant and applies it
consistently across the bin/ scripts. Previously the somatic/germline split was
re-implemented inline in several places with divergent forms (single-column VAF
in filter_cohort.py, three-column elsewhere, plus hardcoded thresholds in
check_contamination.py), which made the somatic and germline sets non-symmetric.
The definition now lives in two shared helpers and every call site uses them. Closes #418.

Changes

• Add somatic_mask / germline_mask to bin/utils_filter.py — the canonical all-3-column
  rule (VAF, vd_VAF, VAF_AM vs a caller-supplied threshold), with NumPy docstrings.
• Refactor bin/filter_cohort.py to use the helpers (somatic flagging goes single-column → all-3-column).
• Refactor bin/check_contamination.py: germline/SNP predicates use the helpers; the hardcoded
  0.2 threshold is replaced by a new --somatic-vaf-boundary CLI option fed from
  params.germline_threshold (wired via modules/local/contamination/main.nf + conf/modules.config);
  NumPy docstrings added to all public functions.
• Add bin/test/test_utils_filter.py — 46 unit tests covering every public function in utils_filter.py.

What to review

• Behaviour changes: filter_cohort somatic flagging is now all-3-column; the
  between-samples contamination threshold moves 0.2 → 0.3 (default germline_threshold).

Testing

pytest bin/test/test_utils_filter.py → 46 passed + ruff clean on all touched files.
The full bin/test/ suite has 3 pre-existing failures + 1 collection error in unrelated files
(test_plot_selectionsideplots.py, test_check_samplesheet.py) that are untouched by this PR.