Train ML model only on simulations

[RemiLehe]

Right now, when experimental data does not agree (e.g., because of incorrect calibration), the ML model (which is trained on both) will be negatively impacted. In addition, this does not allow to interactively try different calibration (see #254)

Proposed solution:

• The ML code should train only on simulation data (similar to: T. Boltz et al., arXiv:2403.03225 (2024))
• In the dashboard, interactively changing calibration would automatically change the position of the ML curves.
• Later: we would add a button to automatically infer the best calibration (using Infer calibration between experiment and simulation quantity, based on user guess #139)