Improve Pandas FFI compatibility

[pbower]

Issue Summary

pd.Series (pandas 3+) and pl.Series expose arrow_c_stream but not arrow_c_array. The existing import_array_stream also crashed on dictionary/categorical types because it reconstructed per-chunk schemas without a dictionary pointer.

Changes

1. src/ffi/arrow_c_ffi.rs - field_from_c_schema (line ~2469)

• Now checks schema.dictionary and returns ArrowType::Dictionary(...) instead of the raw index type when the schema is dictionary-encoded. Previously a dict column with UInt32 indices would return ArrowType::UInt32
• now it correctly returns ArrowType::Dictionary(UInt32).

2. src/ffi/arrow_c_ffi.rs - import_array_stream (line ~2394)

• Now keeps the stream schema alive and passes it directly to import_array_zero_copy. This mirrors the approach used by import_record_batch_stream_with_metadata and fixes the null-pointer crash for dictionary-encoded arrays.
• Also adds the Utf8View -> String dtype fixup that was missing.

3. pyo3/src/ffi/to_rust.rs - array_to_rust (line ~239)

• Adds arrow_c_stream as a fallback between arrow_c_array and _export_to_c. This enables pd.Series and pl.Series to be converted to FieldArray. Multiple chunks are concatenated into a single array.