Fix shatter memory DoS by writing per-leaf results

[ClayWarren]

Motivation

• A change made do_one() to return full per-leaf DataFrames and run() to accumulate them in joined_dfs and pd.concat, allowing unbounded memory growth for large/adversarial inputs.
• The goal is to avoid retaining all per-leaf data in memory while preserving the existing per-leaf processing and total point accounting.

Description

• Updated src/silvimetric/commands/shatter.py so do_one() now returns an int point count and writes each leaf immediately instead of returning a DataFrame.
• do_one() now sorts the per-leaf joined data by ['xi','yi'] and calls write(...) to persist the leaf and return the written point count, and returns 0 for skipped/empty leaves.
• run() no longer accumulates DataFrames; it aggregates per-leaf point counts via point_count from Dask futures or compute(...) results and updates config.point_count accordingly.
• Removed the pd.concat(joined_dfs) path to eliminate unbounded in-memory aggregation while preserving deterministic per-leaf ordering.

Testing

• Ran python -m compileall src/silvimetric/commands/shatter.py which succeeded and verified the modified file compiles.