BUG: TruncatedNormal does not support data imputation

[twiecki]

Describe the issue:

Posterior sampling over a Normal with observed nans works, while the same doesn't work with a TruncatedNormal.

Reproduceable code example:

import pymc as pm

with pm.Model():
    mu1 = pm.TruncatedNormal("mu1", mu=1, sigma=2, lower=0)
    pm.Normal("x1", mu=mu1, sigma=.5, observed=np.array([.1, .2, .5, np.nan, np.nan]))
    pm.sample() # works

    mu2 = pm.TruncatedNormal("mu2", mu=1, sigma=2, lower=0)
    pm.TruncatedNormal("x2", mu=mu2, sigma=.5, lower=0, observed=np.array([.1, .2, .5, np.nan, np.nan]))
    pm.sample() # does not work

Error message:

<details>

SamplingError                             Traceback (most recent call last)
Cell In[155], line 8
      6 mu2 = pm.TruncatedNormal("mu2", mu=1, sigma=2, lower=0)
      7 pm.TruncatedNormal("x2", mu=mu2, sigma=.5, lower=0, observed=np.array([.1, .2, .5, np.nan, np.nan]))
----> 8 pm.sample()

File ~/micromamba/envs/pymc5/lib/python3.11/site-packages/pymc/sampling/mcmc.py:706, in sample(draws, tune, chains, cores, random_seed, progressbar, step, nuts_sampler, initvals, init, jitter_max_retries, n_init, trace, discard_tuned_samples, compute_convergence_checks, keep_warning_stat, return_inferencedata, idata_kwargs, nuts_sampler_kwargs, callback, mp_ctx, model, **kwargs)
    704 ip: Dict[str, np.ndarray]
    705 for ip in initial_points:
--> 706     model.check_start_vals(ip)
    707     _check_start_shape(model, ip)
    709 # Create trace backends for each chain

File ~/micromamba/envs/pymc5/lib/python3.11/site-packages/pymc/model/core.py:1702, in Model.check_start_vals(self, start)
   1699 initial_eval = self.point_logps(point=elem)
   1701 if not all(np.isfinite(v) for v in initial_eval.values()):
-> 1702     raise SamplingError(
   1703         "Initial evaluation of model at starting point failed!\n"
   1704         f"Starting values:\n{elem}\n\n"
   1705         f"Logp initial evaluation results:\n{initial_eval}\n"
   1706         "You can call `model.debug()` for more details."
   1707     )

SamplingError: Initial evaluation of model at starting point failed!
Starting values:
{'mu1_interval__': array(-0.22790012), 'x1_unobserved': array([1.02430783, 0.49541385]), 'mu2_interval__': array(0.51429339), 'x2_unobserved_interval__': array([-inf, -inf])}

Logp initial evaluation results:
{'mu1': -1.48, 'x1_unobserved': -0.74, 'mu2': -0.79, 'x2_unobserved': -inf, 'x1_observed': -2.53, 'x2_observed': -12.71}
You can call `model.debug()` for more details.

PyMC version information:

5.8.0

Context for the issue:

No response