Linearize additive systs around param model; add --forceLinear

[davidwalter2]

Summary

• Scale additive systematics by the param-model value at the linearization point. For systematic_type == \"normal\" with a non-trivial param model (e.g. ABCD / SmoothABCD whose POIs typically sit far from their default), the linearized variation rnorm(poi) * norm + theta * logk did not naturally scale with rnorm, so a ±20% effect defined at the MC nominal became a different relative effect once rnorm moved away from 1. Now logk is pre-multiplied by rnorm_init = param_model.compute(xparamdefault) once at fitter construction. The scaling is a constant so the hot path stays strictly linear in theta (no theta·poi bilinearity). log_normal already had this property, so its tensors are just aliased — no copy.
• Add --forceLinear to force self.is_linear = True regardless of the structural checks. minimize() then takes a single Cholesky / Hessian-CG step from the current x (a Gaussian approximation around that point). Intended for iterative linearized unfolding, where the outer loop re-anchors the linearization point.
• Minor refactor in parsing.py: pull VALID_BIN_BY_BIN_STAT_TYPES directly from rabbit.bbstat.bbstat instead of importing the full fitter module.
• Black formatting on the three tutorial notebooks.

Test plan

• Existing test_tensor.hdf5 (log_normal) and test_tensor_linearized.hdf5 (normal) fits still converge to EDM ≈ 1e-16 with no flag.
• Inline sanity check confirms self.logk scales process 0 by 2.5 and leaves other processes untouched when a test param model with xparamdefault=[2.5] is plugged in.
• --forceLinear emits the info message and takes the Cholesky path on test_tensor_linearized.hdf5.
• Validate end-to-end inside the iterative linearized unfolding loop on a real W/Z analysis input.

🤖 Generated with Claude Code