Warn on silently dropped quadratic/self-interaction terms in DoE formulas

[Copilot]

Formulaic drops non-braced quadratic/self-interaction terms like x**2, x*x, and x:x during parsing, which made DoE model term counts wrong without any signal to the user. This change adds targeted detection and warnings so risky formulas are surfaced early in both strategy utilities and DoE criterion validation.

• AST-based risky pattern detection

  • Added Formulaic AST traversal to detect patterns that are commonly dropped silently:
    • operators outside the supported set for this context (e.g. ** not wrapped in {...})
    • self-interactions (x*x, x:x) that collapse under Formulaic grammar
  • Emits a UserWarning with concrete remediation: wrap intended nonlinear terms in {...}.

• Warnings in both entry points

  • bofire/strategies/doe/utils.py (get_formula_from_string)
  • bofire/data_models/strategies/doe.py (DoEOptimalityCriterion.validate_formula)
  • This closes the silent-failure gap whether formulas come from strategy execution paths or data-model validation paths.

• Focused coverage for warning semantics

  • Added/updated unit tests to assert warnings are raised for risky unbraced terms and not raised for braced equivalents.

from bofire.strategies.doe.utils import get_formula_from_string

# warns: may be silently reduced by Formulaic
get_formula_from_string(model_type="1 + x1 + x1**2")

# no warning: explicit nonlinear intent
get_formula_from_string(model_type="1 + x1 + {x1**2}")