Mask zero points in exp data

[rozyczko]

If a point has a zero value (define zero!), it needs to be masked in the dataset and not used for fitting.
This is to conform to the decision made between the reduction and analysis teams about how to handle zero-intensity points (with undefined varance).

[rozyczko]

masking biases the fit
not clear what the solution is
check other software then - refl1d, genx

[rozyczko]

Currently implemented - masking out for fitting, not for printout.
This is chosen (until told to reimplement) in order to pass the integration tests.

[rozyczko]

@jfkcooper provided this analysis (attached).

It seems using Mighell objective function is probably the best idea, being a least-square method, handling zero values without discarding them and doesn't require additional statistical machinery of MLE (required by Poisson deviance).

Alternatives to Chi squared.docx

[rozyczko]

Current behavior in easyreflectometry.fitting.MultiFitter masks points with zero variance before fitting. This avoids division-by-zero, but discards information and can bias results.

The issue discussion and attached analysis document conclude:

• masking zero points can bias fits,
• Poisson deviance MLE can perform very well but has practical and interpretation drawbacks,
• Mighell's objective function is a good least-squares-compatible compromise for zero-count / zero-variance handling.

Important: In the current codebase, DataSet1D.ye stores variances (σ²), not standard deviations (σ), despite the suggestive field name. This is because load_as_dataset() in measurement.py populates ye from scipp .variances. The existing fit_single_data_set_1d correctly treats ye as variance (weights = 1/sqrt(ye)). All implementation code in this plan must follow the same convention - ye and variances are σ², never to be squared again.

We are going to implement a hybrid solution

• For points with valid variance (variance > 0): standard weighted least squares
  $$\chi^2_i = \frac{(y_i - f_i)^2}{\sigma_i^2}, \quad \sigma_i^2 = \text{variance}_i$$

• For points with invalid/zero variance (variance <= 0): Mighell-like term
  $$\chi^2_{\gamma,i} = \frac{(y_i + \delta_i - f_i)^2}{y_i + 1}, \quad \delta_i = \min(y_i, 1)$$

This follows the Mighell form used in low-count regimes and avoids singular weights at zero variance.

[rozyczko]

Changes to the MultiFitter:

• New enum/string option (internal and public):
  • legacy_mask (current behavior, for strict backward compatibility)
  • mighell (full Mighell transform for all points)
  • hybrid (recommended default for this issue: chi-square for valid variances + Mighell for zero/invalid variances)
  • auto (alias to hybrid for now; future policy hook)

[rozyczko]

branch mighell_transform and corresponding PR: #303