fix(cgan): make density output activation configurable

[SoheylM]

Summary

• add a backward-compatible tanh|sigmoid generator output activation option for cgan_2d and cgan_cnn_2d
• keep tanh as the default so existing W&B checkpoints load with their original convention
• let new density-field runs opt into native [0, 1] sigmoid outputs and document the flag

Fixes #69

Tests

• conda run -n engibench pytest -q
• conda run -n engibench ruff check engiopt/cgan_2d/cgan_2d.py engiopt/cgan_2d/evaluate_cgan_2d.py engiopt/cgan_cnn_2d/cgan_cnn_2d.py engiopt/cgan_cnn_2d/evaluate_cgan_cnn_2d.py tests/test_cgan_output_activation.py
• conda run -n engibench ruff format --check engiopt/cgan_2d/cgan_2d.py engiopt/cgan_2d/evaluate_cgan_2d.py engiopt/cgan_cnn_2d/cgan_cnn_2d.py engiopt/cgan_cnn_2d/evaluate_cgan_cnn_2d.py tests/test_cgan_output_activation.py

[SoheylM]

Adding the qualitative Beams2D sanity-check figure for this cGAN output-activation change.

This is a short local smoke test, not a final benchmark: cgan_2d was trained for 50 epochs on 2048 Beams2D training samples and compared against optimized test references under the same conditions. The main motivation for this PR shows up clearly: the legacy tanh head still emits invalid negative raw densities (raw_neg=27.7% here), while the opt-in sigmoid head emits native [0, 1] densities (raw_neg=0.0%) and modestly improves this quick reference MAE (0.191 vs 0.210).