Improve and standardize the cuPDLPx Python API

[LucasBoTang]

Summary

This PR hardens and standardizes the Python interface for cuPDLPx. It fixes binding/API correctness issues, strengthens validation, and makes the public API more consistent with common optimization solver conventions.

Highlights

• Fixed correctness issues in the Python binding layer, including a solve_once result-lifetime leak and an incorrect status-code mapping.
• Improved binding memory safety with RAII for native problem and result objects.
• Strengthened validation for model data and bounds, warm starts, solver parameters, and sparse matrix inputs.
• Standardized solver status reporting:
  • Model.Status is now an integer status code.
  • Model.StatusName provides the human-readable status.
  • The redundant Model.StatusCode property was removed.
• Improved model and parameter ergonomics:
  • optimize() now returns the model instance.
  • resetParams() restores default parameter values.
  • Params supports attribute access, item access, and lightweight mapping-style helpers.
• Added support for SciPy sparse arrays (csr_array and coo_array) alongside legacy sparse matrices.
• Updated documentation and tests, and removed obsolete code and pytest configuration.

Breaking changes

• Model.Status is now an int; use Model.StatusName for the textual status.
• Model.StatusCode has been removed.
• setWarmStart() now raises ValueError when provided vectors have incorrect dimensions.
• Invalid model data, bounds, sparse matrices, and parameters are rejected earlier and may now raise Python exceptions before entering the solver.

Testing

• Rebuilt the Python extension in WSL.

• Passed the Python API tests and full Python test suite:

  pytest test/test_api_surface.py
  pytest test