Close leaked CPS HDF stores

[MaxGhenis]

Summary

• close raw CPS HDF stores after previous-year income preprocessing and after the auto-loan/net-worth preprocessing path
• close the initial raw CPS store as soon as the base CPS entity tables are materialized
• add a focused regression test that verifies the previous-year CPS handles are closed

Testing

• UV_PYTHON=3.14 uv run pytest tests/unit/datasets/test_cps_file_handles.py tests/unit/datasets/test_cps_takeup.py tests/unit/test_extended_cps.py -q
• python3 -m py_compile policyengine_us_data/datasets/cps/cps.py tests/unit/datasets/test_cps_file_handles.py
• UV_PYTHON=3.14 uv run ruff format --check policyengine_us_data/datasets/cps/cps.py tests/unit/datasets/test_cps_file_handles.py

[baogorek]

Looks good to me. Completely optional, but here's an argument for context managers over try/finally:

For a two-line body the try/finally is a lot of ceremony. pd.HDFStore supports with, so:

  with self.raw_cps(require=True).load() as raw_data:                                                         
      raw_person = raw_data["person"]
      cps["is_married"] = raw_person.A_MARITL.isin([1, 2]).values                                             
                                                                                                              
  Same behavior, less noise. I'd recommend context managers throughout if the PR author wants to close these  
  handles. 

[baogorek]

the try/finally is arguably overkill as the list comprehension over a fixed ENTITIES constant is unlikely to fail. A context manager (with) would be cleaner if the HDFStore supports it (and Pandas HDFStore does)