goldenmatch v3.1.0

Changed

• polars is now OPTIONAL (pip install 'goldenmatch[polars]'). The engine
  is Arrow-native end to end: ingest, prep (incl. the goldencheck quality scan
  on its Arrow surface and the goldenflow transform adapter), matchkey
  precompute, blocking, exact matching, scoring (classic + bucket backends
  with the Rust kernels), clustering, golden survivorship (fused kernel +
  seam-native oracle), memory corrections, identity resolution, lineage, and
  file outputs (native parquet). A new zero-polars gate
  (tests/test_zero_polars_gate.py) proves a full dedupe with polars imports
  blocked.
• With polars installed, the wall-optimization paths (golden fast columnar,
  vectorized survivorship, the vectorized pair-score join) light up
  automatically and behavior is byte-identical to 3.0.x.
• GOLDENMATCH_FRAME=polars (the classic opt-out lane) now requires the
  [polars] extra and raises a clear error without it.
• The Frame lane now accepts EVERY feature class (validation, outputs,
  lineage, identity, memory, auto-suggest, postflight, probabilistic EM,
  NE-on-exact, throughput, rerank, LLM, semantic blocking, domain
  extraction, adaptive golden) -- the eligibility predicate has no feature
  declines left.

Fixed

• most_recent golden rules on date32/time32 columns crashed on the arrow
  lane (pc.cast has no direct 32-bit temporal -> int64 kernel).