Nutpie fails to sample when RV changes shape following `with_data`

import pymc as pm
import nutpie

with pm.Model() as model:
    x = pm.Data('x', [0,1,2])
    beta = pm.Normal('beta', mu=0, sigma=1)
    y = pm.Normal("y", mu=beta * x, sigma=1)   # fine if it's observed and we also change it's data

compiled_model = nutpie.compile_pymc_model(model)

compiled_model_set = compiled_model.with_data(x=[1,2])
nutpie.sample(compiled_model_set, chains=1)
# RuntimeError: All initialization points failed

# Caused by:
#     Logp function returned error: ErrorCode(1)

Start by running the reproducer using nutpie.compile_pymc_model, compiled_model.with_data, and nutpie.sample with the changed x shape. Trace how with_data handles the shape change and how sampling initialization reports the log-probability error. Done means the shown model samples successfully after x changes from three to two values, with regression coverage for this case.