Add scaling factor

[lmanan]

Normalize features and costs by a common scaling factor before solving

The soft margin loss optimizer can suffer from numerical instability when features and cost coefficients operate at very different scales. This PR introduces a scaling factor to normalize them before the MIP solver runs.

Changes:

• LinearCosts: add set_scaling_factor / get_scaling_factor methods
• HammingCosts: set scaling factor to the number of active variables (i.e. sum(mask) if a mask is provided, else n)
• SoftMarginLoss: divide _b, _g, and _features by the scaling factor before optimization, and use the scaled features when computing the gradient
• SoftMarginLoss.value_and_gradient: explicitly add the constant term a + b back to the solver's return value, since get_value() excludes it