tests: add xfail tests for float lapse_dist bugs in _apply_lapse_model

[cpaniaguam]

@AlexanderFengler @digicosmos86 Would it be sensible to address this?

Summary

Adds two xfail(strict=True) tests that document two latent bugs in _apply_lapse_model (in dist.py). The tests are currently expected to fail and will flip to passing once a fix is added.

---

Background

For choice-only models (e.g. softmax_inv_temperature_2), _check_lapse sets self.lapse = 1 / n_choices — a plain float. This float is stored as HSSMRV._lapse and passed to _apply_lapse_model(lapse_dist=...) during prior/posterior predictive sampling. The existing code assumes lapse_dist is always a bmb.Prior object, so two things go wrong.

---

Bug 1 — AttributeError on lapse_dist.args

_apply_lapse_model always executes:

lapse_rt = pm.draw(
    get_distribution_from_prior(lapse_dist).dist(**lapse_dist.args), ...
)

When lapse_dist is 0.5 (a float), lapse_dist.args raises:

AttributeError: 'float' object has no attribute 'args'

Trigger: any call to model.sample_prior_predictive() or model.sample_posterior_predictive() on a choice-only model that has p_outlier != 0.

---

Bug 2 — 2-column replace mask for a 1-column output

Even if Bug 1 were somehow bypassed, the replacement mask is always built as:

replace_mask = np.stack([replace, replace], axis=-1)  # always 2 columns

Choice-only simulation output has only 1 column (response; no RT). The 2-column mask either silently corrupts the array via np.putmask broadcasting or raises a shape error depending on input dimensions.

Trigger: same as Bug 1.

---

Current status

Both bugs are latent on the add-choice-only-rlssm-model-to-registry branch. For RLSSM models ssms_rng_fn raises before reaching _apply_lapse_model (ssm_simulators does not support 2AB_RescorlaWagner_Softmax), so neither bug is reachable through the RLSSM path today. They would be reachable for any other choice-only model supported by ssm_simulators (e.g. softmax_inv_temperature_2) the moment a user passes p_outlier != 0.

---

Fix (tracked separately)

Add an isinstance(lapse_dist, float) branch in _apply_lapse_model that:

• skips RT sampling (there is no RT column)
• samples replacement responses uniformly over choices
• handles both 1-D (n_obs,) and 2-D (n_obs, 1) sims_out

The type annotation of make_hssm_rv's lapse parameter should also be updated from bmb.Prior | None to bmb.Prior | float | None.