feat: preserve typed Python data source failures

[Robin-Everaars]

Summary

• Add an opt-in Python DataSource failure marker with exact terminal and transient values.
• Preserve the finite category through DataFusion and remote task transport while discarding classified Python messages and tracebacks.
• Map terminal failures to AnalysisException and transient failures to SparkRuntimeException.
• Keep unknown or absent marker values on the existing Python exception path.
• Document the contract and cover direct readers, generator readers and generic Python UDF isolation.

Upstream context

• lakehq/sail#2242 tracks the larger need for structured internal errors translated to Spark conditions. This pull request is intentionally narrower: application opt-in at the Python DataSource boundary rather than a general DataFusion or Arrow error taxonomy.
• lakehq/sail#2267 is developing a first-party Python Kafka source. This protocol remains connector-agnostic and does not define Kafka retry policy.

Verification

• cargo test
• cargo clippy -p sail-common-datafusion -p sail-python-udf -p sail-data-source -p sail-spark-connect --all-targets --all-features -- -D warnings
• cargo fmt --check
• Wheel build with maturin
• Focused installed-wheel Spark Connect tests: 2 passed
• Existing unclassified DataSource exception check
• Exact-head GitHub Actions run 30447886867: Rust build/tests, Python tests for four supported Spark versions and docs passed
• Exact-head Rust and Python lint runs passed

Live Spark, Ibis and catalog integration jobs were not selected by the exact-head CI run.