Improve Rust reader throughput and resilience

[sebastian-flyr]

Summary

• Increase the Rust BigQuery Storage reader decode limit to handle larger Arrow ReadRows messages.
• Use one BigQuery Storage client/channel per read stream while sharing a fetched auth token.
• Batch currently available Rust-decoded record batches per Python iterator step and pass batch_size into the Rust reader.
• Reduce Rust stream fanout from worker_count * 3 to worker_count and retry transient ReadRows stream failures from the last decoded offset.

Test plan

• cargo check --manifest-path /tmp/pyarrow-bigquery/rust/Cargo.toml
• uv run --group dev ruff check .
• Full BigQuery Rust load from consuming project: 695678 rows x 35 columns in 44.37s
• Five repeated full BigQuery Rust loads from consuming project: 24.43s, 28.38s, 29.37s, 26.19s, 28.49s

Note: uv run --group dev pytest currently fails during collection because the package test venv cannot import pyarrow.bigquery under the installed pyarrow package layout. This appears to be an existing packaging/test-environment issue independent of these reader changes.