fix(s3): prevent per-shard client and cache growth

[durson]

• Reuse aiobotocore clients per process and S3 configuration instead of creating a retained client context for every shard. Reset inherited client state in DataLoader workers so clients remain tied to the worker event loop.

• Bound FileReader's async range cache to a configurable per-reader memory budget and expose presigned retry count for tail-latency tuning. Preserve the previous default of four total presigned request attempts.

• Bound presigned S3 request timeouts to avoid DataLoader stalls

Added env vars:

• WSDS_ASYNC_CACHE_MAX_BYTES (max cache size per thread)
• WSDS_S3_PRESIGNED_RETRIES (configurable number of re-tries)
• WSDS_S3_PRESIGNED_TIMEOUT_SECONDS
• WSDS_S3_PRESIGNED_CONNECT_TIMEOUT_SECONDS