BEHAVIOR: use subtraction in UnbinnedNLL

[redeboer]

Closes #565

⚙️ Enhancements

• UnbinnedNLL.__call__ now computes N * log(integral) - sum(log(data)) instead of -sum(log(data / integral)), which is the formula its own docstring already documents. This drops the full normalized-likelihood temporary: roughly 1.5x faster at formula level on NumPy, and ~1.15x on the full estimator call, where the intensity evaluation dominates.

❗ Behavioral changes

• The two formulations are only equivalent when the normalization integral and all data intensities are strictly positive. Where the division form stayed finite for a positive ratio of two negative numbers, the subtraction form takes the two logarithms separately and returns NaN.
• Concretely: a model with unconstrained signed coefficients (a very common way to write a mixture) can be pushed into negative intensities by an unbounded optimizer, and the fit now returns NaN instead of converging on a physically meaningless minimum. This is not hypothetical — it broke test_optimize_all_parameters for all four back-ends under ScipyMinimizer.
• If you hit this, make the model non-negative by construction rather than relying on the old behavior. Squaring the coefficients is enough, and is what the test and benchmark models now do:

  expression = a**2 * gaussian(x, mu1, sigma1) + b**2 * gaussian(x, mu2, sigma2)
  parameter_defaults = {a: np.sqrt(0.15), b: np.sqrt(0.05)}

  Fitted values are then square roots of the old ones, so any hard-coded reference values need regenerating.

🖱️ Developer experience

• New benchmarks/unbinned_nll.py compares the two formulations over NumPy, Numba, JAX, and TensorFlow. Each back-end is timed with its own warm-up and synchronisation (block_until_ready() for JAX, .numpy() for TensorFlow, a compilation call for Numba), so that the numbers are actually comparable, and every benchmark asserts its result against a NumPy reference.
• The same file also benchmarks the full UnbinnedNLL estimator per back-end, so that the formula-level difference can be weighed against the intensity evaluation that dominates a real fit.
• The Numba implementations are written as explicit prange loops under njit(parallel=True) instead of as array expressions, since that is where Numba is competitive at all.
• poe benchmark now takes optional positional paths, so that a single file can be run in isolation instead of the whole benchmarks/ directory:

  poe benchmark benchmarks/unbinned_nll.py

• benchmarks/expression.py and the fit test model square their mixture coefficients, so that they are non-negative by construction.

🔨 Maintenance

• Renamed the private back-end attributes __mean_function, __sum_function, and __log_function to __mean, __sum, and __log, and the local bare_intensities to data_intensities.

Squash commit messages

* DX: allow passing paths to `poe benchmark`
* DX: benchmark `UnbinnedNLL` implementations per back-end
* DX: enforce positive intensities in test and benchmark models
* MAINT: rename private back-end attributes in `UnbinnedNLL`