Fix file descriptor leaks in wsds

[tlebryk]

SUMMARY: pyarrow's RecordBatchFileWriter.close() doesn't release the fd — only GC does. Force gc.collect() after close and add explicit close() methods to WSShard/WSDataset so file handles are released.

• RecordBatchFileWriter.close() flushes metadata but keeps the fd open until the Python object is garbage-collected. On container reuse, volume.reload() fails with "there are open files preventing the operation" because the previous invocation's writer fds are still held.

• WSDataset caches WSShard objects in _open_shards, each holding a memory-mapped file via pa.memory_map().

• _build_key_iter created a temporary WSDataset (with open shard handles) via a lazy generator that was never closed, leaking handles until GC.

Changes:

• ws_sink.py: gc.collect() after RecordBatchFileWriter.close() to force fd release; close writer on error path too; _build_key_iter eagerly collects keys and closes the temp dataset; catch KeyError for race condition where sibling column dir exists but has no committed shards yet
• ws_shard.py: Add close() method; keep reference to _source_file so we can explicitly close the pyarrow NativeFile (memory_map/OSFile)
• ws_dataset.py: Add close() that closes all cached shards and linked datasets