v0.2.18

• Add an experimental DuckDBBaseConverter to convert large Parquet-based datasets
  without loading them into memory. Its output matches the default converter
  (geometry handling, Hilbert order, data types, metadata, file packaging).
  duckdb is a new dependency.
• Converters record the collection id in the collection metadata and no longer add a
  constant collection column. Previously files without constant columns were written
  without any collection id.
• Converters drop rows that can never validate (missing required values, empty or
  missing geometries). Missing required values are dropped only up to the new
  max_dropped_share (default 1%), above it the conversion fails.
• Converters fail when both sources and variants are declared.
• Converters warn when no column is mapped to id and when the id column is not unique.
• Converters load all schemas upfront with retries, so a temporary network issue
  no longer kills a long conversion at the very end.
• Send User-Agent: vecorel-cli on HTTP downloads instead of fsspec's default. Servers that
  reject the default answer 403, which surfaced as FileNotFoundError naming only the URL and
  read as a dead source.