Demonstrates Dispatch with all three batching strategies, streaming, and fault isolation via per-worker process boundaries.

Generates 50 numeric datasets of varying sizes (small, medium, large) plus 5 "poison pill" datasets that contain invalid data (nil, NaN, empty). Dispatches them to a Python batch_stats module using:

1. FixedBatch -- fixed chunks of 10 items
2. WeightedBatch -- batches capped at 500 total values
3. KeyPartition -- one batch per dataset size category
4. Streaming -- lazy stream with as_completed ordering

After all strategy runs, queries each worker for its accumulated running statistics (Welford's online mean/variance) to show that per-process state remains internally consistent -- something impossible under free-threaded Python where concurrent threads corrupt shared accumulators.

Run with Slither.Examples.BatchStats.StatsDemo.run_demo/0.