Fix RLCD stage-one setup and dependency handling

[kunwuz]

Summary

• use pure-Python GES as RLCD's only supported sample-data stage-1 method
• pass stage1_ges_sparsity to GES through the effective lambda_value argument
• replace the optional communities dependency with NetworkX for maximal-clique enumeration
• remove the unreachable Java stage-1 integration that referenced files absent from causal-learn
• validate stage-1 choices and document the supported behavior
• add regression coverage for the public default path, parameter forwarding, method validation, clique partitioning, and latent discovery

Background

The NumPy 2 scalar-conversion failure in the released 0.1.4.7 package is real, but it was already fixed on main by #263. The remaining RLCD stage-1 issues were still present on current main: an alternate backend referenced files absent from causal-learn, stage1_ges_sparsity was silently ignored by GES, and importing RLCD required an optional package that a base installation did not provide.

The GES sparsity default changes from 2 to 0.5 to preserve the value that GES previously used after ignoring the obsolete lambda key.

Testing

• pytest -q tests/TestRLCD.py: 5 passed with NumPy 1.26.4
• pytest -q tests/TestRLCD.py: 5 passed with NumPy 2.4.4
• the documented five-indicator latent example passes with the GES stage
• fresh wheel smoke test passed with NumPy 2.5.1 and no communities installation
• RLCD package compilation passed
• git diff --check passed

The existing GES accuracy assertion also fails on untouched main; this change does not modify GES scoring code.