Fix Snappy decompressor buffer corruption and add ClickBench tests

[unexge]

Summary

Fixes buffer corruption issues in the Snappy decompressor that occurred when processing large files with copy operations.

Changes

Snappy Decompressor Fix

• Root cause: The decompressor was using the writer's buffer for copy back-references, but this buffer could be invalidated/flushed between writes, causing corruption
• Fix: Introduced a dedicated circular window buffer (128KB) that maintains the last bytes for copy operations, independent of the writer's state
• Added proper bounds checking for copy offsets (copy.offset > d.total_written or copy.offset == 0)
• Added EOF handling when remaining == 0 in both literal and copy states
• Added buffer length assertion in init()

ClickBench Conformance Tests

• Added download script for ClickBench partitioned Parquet files (~450MB total, CI only)
• Added conformance test that reads all 3 ClickBench files (105 columns each)
• Updated README with ClickBench dataset documentation

Testing

• All existing Snappy tests pass
• ClickBench tests exercise the fix with real-world large Snappy-compressed data