feat: add configurable DuckDB memory limit via DUCKDB_MEMORY_LIMIT env var

[CarstVaartjes]

Summary

• Add DUCKDB_MEMORY_LIMIT environment variable support to cap DuckDB memory per query connection
• When set (e.g. DUCKDB_MEMORY_LIMIT=2GB), DuckDB spills to temp storage instead of allocating unbounded memory
• Prevents OOM on containers with limited memory (e.g. ECS tasks running multiple Gunicorn workers)

Context

DQE reader runs 3 Gunicorn workers in a 7.5GB ECS container. DuckDB's default behavior allocates memory without bounds, causing OOM on large shard aggregations. With DUCKDB_MEMORY_LIMIT=2GB, each worker caps at 2GB (3×2GB=6GB), leaving headroom for Python/OS.

Benchmark (345M-row KCI shard, high-cardinality query)

Limit	Time	RSS delta	Result
None (default)	7.93s	+2,713 MB	256K rows
2GB	6.71s	+540 MB	256K rows
512MB	3.17s	OOM	Failed

No performance penalty — bounded memory is actually faster due to less GC pressure.

Changes

• parquery/aggregate_duckdb.py: Read DUCKDB_MEMORY_LIMIT env var, pass as config to duckdb.connect()
• parquery/__init__.py: Version bump to 2.0.7
• RELEASE_NOTES.md: Added 2.0.7 entry

Test plan

• All 76 existing tests pass
• Verified with real 345M-row parquet file at various limits
• Deploy to DQE reader with DUCKDB_MEMORY_LIMIT=2GB env var