New CTable dictionary spec, nested columns and richer Arrow/Parquet interoperability

[FrancescAlted]

This PR extends CTable with richer Arrow/Parquet interoperability, nested-column support, dictionary-encoded string columns, faster table opening, and lower blosc2 import overhead.

Main additions:

• Add first-class CTable dictionary/categorical string columns via blosc2.dictionary().
  • Stored as compact int32 codes plus a persisted global string dictionary.
  • Arrow dictionary columns are preserved by default during Parquet/Arrow import.
  • Added equality, isin(), null handling, Arrow export, persistence, and CLI roundtrip support.
  • --decode-dictionaries is available in the Parquet CLI to opt out and import as vlstring.
• Add nested Arrow/Parquet struct support for CTable.
  • Top-level Arrow/Parquet structs are recursively flattened into dotted CTable leaf columns, e.g. trip.begin.lon.
  • Leaf columns remain independent compressed arrays for fast analytics.
  • Row reads reconstruct the original nested dict shape.
  • Dotted leaves work with getitem, attribute access, where(), select(), indexes, aggregates, and Parquet roundtrips.
  • Literal ., /, and \ in field names are escaped and stored safely.
• Add support for separating nested list<struct<...>> Parquet layouts.
  • Especially useful for Awkward-style unnamed-root list datasets.
  • CTable.from_parquet(..., separate_nested_cols=True) is now the default.
  • Each list element becomes a CTable row, and struct leaves become regular nested columns.
  • max_rows limits flattened CTable rows in this mode.
• Improve parquet_to_blosc2 CLI.
  • Adds import/export/roundtrip workflows.
  • Adds progress reporting with ETA.
  • Adds nested-column controls via --separate-nested-cols / --no-separate-nested-cols.
  • Defaults list imports to Arrow serialization for better performance on deeply nested data.
  • Adds dictionary preservation/decode controls.
  • Adds --max-rows, timestamp handling improvements, memory reporting, profiling, and roundtrip comparison helpers.
• Improve CTable persistence/open performance.
  • Adds lazy column metadata/data loading for faster CTable.open().
  • Makes CTable.nrows lazy.
  • Avoids unnecessary store-extension validation on open.
  • Preserves nested schema metadata and empty-root Arrow names across save/load/open/export paths.
• Improve base blosc2 import time.
  • Lazy-imports numba and requests.
  • Avoids eager NumPy callable scans that triggered expensive lazy NumPy submodule imports.
  • Local profiling reduced import blosc2 from roughly ~225 ms under importtime to ~82–86 ms.
• Add documentation, plans, benchmarks, and tests.
  • New CTable docs for Parquet interoperability, nested fields, dictionary columns, null policy, indexing, and querying.
  • New tests for dictionary columns, nested access/storage, nested metadata, Parquet interop, getitem access, and BatchArray metadata behavior.