feat(pull): add pull-based streams adapter

[tryangul]

What

PullBasedAdapter — new streams adapter backed by the arroyo pull-based runtime.

Depends on arroyo#559.

How

• PullBasedAdapter — drop-in StreamAdapter replacement for RustArroyoAdapter. Translates DSL steps (HeadersFilter, Batch, Map, GCSSink, etc.) into pull-based operators. Registered as "pull" in loader.py.
• PullConsumer — the runtime entry point. Rebuilds fresh stages on each rebalance (matching the push model's destroy/recreate pattern). Parameterless run() with SIGINT/SIGTERM handling.
• PipelineValue — type-erased payload (Raw/Rust/Python) so the adapter can compose stages dynamically from the DSL without compile-time generics for every pipeline shape.
• PyCallableStage — bridges Rust and Python. Wraps input in a streams Message object before calling so existing callables (batch_msg_parser, ItemsSpanProcessor, serialize_to_parquet) work unchanged.
• GcsClient — injectable GCS upload client, separated from sink handler for testability.

Why

Steel thread for the pull-based runtime — proves the arroyo primitives work end-to-end with the streams DSL and real pipeline callables (protobuf parsing, parquet serialization).

Tests (11)

• 6 Rust — stage unit tests, consumer e2e, graceful shutdown
• 5 Python — adapter translation, e2e with Message wrapping, real BatchParser (protobuf via sentry_kafka_schemas) + ParquetSerializer (polars) through the pipeline

Notes

• Stages rebuilt fresh on each rebalance — no stale batch state
• No test infrastructure in production code — Python e2e tests use a capturing Map step
• Concurrent GCS uploads not yet supported — straightforward via .apply_concurrent()

Next steps

• Deploy to SBC items_span to benchmark
• Batch timeout flush
• GCS upload retry (backon)

STREAM-1668