Parameter identifiability and uncertainty reports #25

Context

SVI parameters can be poorly identified while the fitted smile is excellent: near-indistinguishable smiles from very different (a,b,rho,m,sigma). Stable fitted IV is not stable fitted parameters -- signals built on parameter changes can be pure optimizer noise.

Proposal

Per-fit identifiability tooling:

• fit.jacobian() (dw/dtheta at the quotes), fit.condition_number()
• fit.parameter_uncertainty() -- covariance from the Gauss-Newton approximation at the optimum (residual-scaled), reported as per-parameter standard errors and correlations
• fit.identifiability_report() -- a formatted block flagging well/poorly identified directions (e.g. sigma-b correlation, wing under-determination on narrow strike ranges)

Builds on the analytic derivative kernels; ties into the fit-report wrapper.

Acceptance criteria

• Narrow-strike fixture flags wing parameters as poorly identified while iv RMSE is tiny
• Uncertainties shrink as the strike range widens
• Works for the SVI family analytically; FD fallback elsewhere