goldenmatch v3.2.0

Added

• N-level probabilistic comparison fields (level_thresholds): probabilistic
  matchkey fields accept explicit per-level similarity cutoffs (descending,
  len == levels - 1; a pair's level = the count of thresholds its similarity
  satisfies), generalizing the fixed 2/3-level agree/partial/disagree banding.
  Scalar and vectorized scoring paths both honor them; the fused-match path
  declines level_thresholds matchkeys and falls back.
• Native N-level scoring (goldenmatch-native 0.1.14): the native Rust FS
  kernel scores custom level_thresholds banding natively — byte-identical to
  the pure-Python _levels_from_similarity semantics — via an optional
  per-field level_thresholds kwarg on score_block_pairs_fs. Capability is
  detected through the kernel's FS_SUPPORTS_LEVEL_THRESHOLDS const, so older
  wheels keep the automatic pure-Python fallback (no behavior change).
• Splink config converter: from_splink() (top-level export) converts a
  Splink settings or trained-model JSON (dict or path) into a validated
  GoldenMatch config plus a ConversionReport of lossy findings; trained m/u
  probabilities import as an EMResult so no re-training is needed.
  strict=True raises on any lossy finding. Surfaced as the
  goldenmatch import-splink SETTINGS.json -o CONFIG.yaml [--model-out MODEL.json]
  CLI command and the convert_splink_config MCP tool (78 tools total).
• Splink-conversion parity gate (scripts/bench_er_headtohead/run_converted_splink.py):
  a converted config must land within F1 0.05 of native Splink on the shared
  evaluator; measured splink_f1=0.9964 vs converted_gm_f1=0.9761
  (delta 0.0203) on synthetic_person (5K).