axisnorm, axisexp, axisbernstein poi models

[pmlugato]

N.B. This is a draft as it will follow #130, and conform to new naming scheme and file structure introduced there.

Summary

Adds three new POI model classes (AxisNormModel, AxisExpModel, AxisBernsteinModel) for per-cell
signal normalization and exponential or Bernstein polynomial (preliminary just 1st order) background parameterization, enables composing multiple POI models on a single fit via repeated --poiModel flags,
and fixes a bug in CompositePOIModel that froze parameters of sub-models with
unconstrained POIs.

Changes

poi_model.py

AxisNormModel: new class. Assigns one independent normalization POI per
(process, cell) where cells are defined by a caller-specified set of axes
looked up from indata.channel_info. POI = x² reparameterization enforces
non-negativity. All other channels and processes are left at 1.0.

AxisExpModel: new class. Assigns per-cell lnAmpl and per-group slope
POIs producing rnorm = exp(lnAmpl + slope·x) across the shape axis. Both
parameters are unconstrained reals (allowNegativePOI=True); slope=0 (flat
background) is an interior point so the Hessian is non-degenerate there. Slope
can optionally be shared across a coarser axis grouping via an optional 5th CLI
argument slope_axes, reducing parameters and stabilizing the Hessian when
per-cell slopes are under-constrained.

CompositePOIModel bug fix: previously hardcoded allowNegativePOI=False,
causing the fitter to apply x² to the entire combined parameter vector. This
silently froze sub-models with allowNegativePOI=True at initialization (zero
gradient via chain rule at x=0). Fixed by setting allowNegativePOI=True on
the composite and applying the x² transform per-sub-model inside compute()
based on each sub-model's own flag. Also corrects is_linear to require all
sub-models to be linear rather than inheriting the composite flag.

AxisBernsteinModel: new class. Assigns two non-negative POIs (c₀, c₁)
per (process, cell) and produces rnorm(x_m) = c₀·(1−x_m) + c₁·x_m (first-
order Bernstein polynomial) where x_m ∈ [0,1] is the normalized mass bin
center. Non-negativity is enforced via the x² reparameterization
(allowNegativePOI=False); default c₀=c₁=1 gives a flat unit background at
an interior point in parameter space. Structurally analogous to AxisExpModel
but bounded and linear, with no slope-grouping option.

helpers.py: registers AxisNormModel and AxisExpModel in the
built-in model loader so they resolve without a dotted module path.

parsing.py

--poiModel changed from a single nargs="+" argument to action="append" nargs="+", allowing the flag to be repeated. Multiple models are automatically
composed via CompositePOIModel in rabbit_fit.py.

rabbit_fit.py

Updated model instantiation to iterate over all --poiModel specs and wrap
them in CompositePOIModel when more than one is provided.

Usage example

# Exponential background with slope shared per eta bin
rabbit_fit.py tensor.hdf5 -o results/ \
  --poiModel AxisNormModel <channel> signal   pt,eta,charge \
  --poiModel AxisExpModel  <channel> bkgExp   mass pt,eta,charge eta

# First-order Bernstein background, one (c₀,c₁) per cell
rabbit_fit.py tensor.hdf5 -o results/ \
  --poiModel AxisNormModel    <channel> signal     pt,eta,charge \
  --poiModel AxisBernsteinModel <channel> background mass pt,eta,charge