perf: add Cython metadata parser using BytesIOReader (100's-1000's of ns improvements, x1.5-4 speedup)

[mykaul]

Summary

Add cassandra/metadata_parser.pyx with zero-copy metadata parsing using BytesIOReader, eliminating per-read bytes allocation that dominates recv_results_metadata and read_type cost in the Cython path.

Changes

• New file: cassandra/metadata_parser.pyx — Cython cdef inline functions (read_int_br, read_short_br, read_string_br, read_type_br) operating on BytesIOReader instead of Python BytesIO. A factory function make_recv_results_metadata() returns a closure that captures the type-code map and type objects once.

• Modified: cassandra/row_parser.pyx — Creates a single BytesIOReader from the full remaining buffer and reuses it for both metadata parsing and row parsing (previously: Python BytesIO for metadata, then a separate BytesIOReader for rows). Error recovery path correctly saves/restores reader position.

• Modified: cassandra/protocol.py — cython_protocol_handler() imports the new metadata parser and passes it to make_recv_results_rows().

Benchmarks

All measurements: taskset -c 0, Python 3.14.3, Cython compiled, origin/master vs this PR.

recv_results_metadata only (isolated)

Scenario	Before	After	Speedup
3 simple cols	3,022 ns	582 ns	5.2x
10 simple cols	6,082 ns	1,265 ns	4.8x
50 simple cols	26,324 ns	4,882 ns	5.4x
NO_METADATA	541 ns	215 ns	2.5x
10 cols w/ collections	28,163 ns	21,832 ns	1.3x

Collections are limited by apply_parameters cost (addressed separately in #794).

recv_results_rows end-to-end (old Cython vs new Cython)

Scenario	origin/master	This PR	Speedup
3 cols, 0 rows	5,661 ns	2,687 ns	2.04x
10 cols, 0 rows	11,068 ns	5,652 ns	1.96x
50 cols, 0 rows	44,904 ns	20,569 ns	2.18x
10 cols, 10 rows	29,590 ns	24,759 ns	1.20x
10 cols, 100 rows	181,053 ns	173,446 ns	1.04x
10 cols, 1000 rows	1,798,441 ns	1,821,257 ns	1.00x
NO_METADATA, 0 rows	5,733 ns	5,075 ns	1.13x

The improvement is entirely in metadata parsing (~2x for 0-row cases). As row count increases, metadata cost becomes a smaller fraction of total time, so the speedup dilutes. The 1000-row case is dominated by row deserialization and shows no change.

Tests

231 passed, 1 skipped — identical to baseline.

Notes

• The prepared path (recv_prepared_metadata) is unchanged — it still uses Python BytesIO inherited from ResultMessage. Prepared statements are parsed once per PREPARE, not on the hot path.
• setup.py automatically picks up the new .pyx file via the existing cassandra/*.pyx glob.
• Flag constants are duplicated as compile-time DEF in the .pyx file (documented with a sync warning comment).

[mykaul]

I'm slowly but surely converting the Python driver to a C driver.... Unsure about this direction. :-/