Replace assert with proper validation in channel_to_kraus()

[ddri]

Summary

• Replace assert statements with proper if/raise ValueError
• Add descriptive error messages with diagnostic values
• Handle numerical noise gracefully
• Add complete docstring

Problem

The function used assertions for critical validation:

assert np.allclose(evals.imag, 0.0)  # FIXME exception
assert np.all(evals.real >= 0.0)  # FIXME exception

Running python -O disables assertions, causing invalid channels to silently produce garbage Kraus operators with NaN values (from sqrt of negative numbers).

Demonstrated with -O flag before fix:

RuntimeWarning: invalid value encountered in sqrt
  values = np.sqrt(evals.real)

Fix

if not np.allclose(evals.imag, 0.0):
    max_imag = np.max(np.abs(evals.imag))
    raise ValueError(
        f"Choi matrix has non-real eigenvalues (max imaginary: {max_imag:.2e}). "
        "Channel may not be Hermitian-preserving."
    )

if np.any(evals.real < -1e-9):
    min_eval = np.min(evals.real)
    raise ValueError(
        f"Choi matrix has negative eigenvalues (min: {min_eval:.2e}). "
        "Channel may not be completely positive."
    )

# Clamp small negative values from numerical noise to zero
values = np.sqrt(np.maximum(evals.real, 0.0))

Test plan

• Valid channels still convert correctly
• Invalid channels raise ValueError (not AssertionError)
• Validation works even with python -O flag
• All existing channels_test.py tests pass (32 tests)

Fixes #131