aDDM model cartoon: condition on observed fixations + honor continuation policy (currently generic-SSM re-simulation)

Summary

hssm.plotting.plot_model_cartoon produces a valid-looking cartoon for the aDDM, but its predictive is decoupled from the aDDM's fixation covariates and continuation policy in three ways. The cartoon reuses the generic-SSM path, which assumes "re-simulate at posterior θ" fully specifies the predictive — true for covariate-free SSMs, but the aDDM adds per-trial fixation covariates (r1, r2, flag, sacc_array, d, sigma) and a posterior-predictive continuation policy that the cartoon has no way to thread.

Follow-up to #1032 (cartoon) and #1035 (fixation continuation). A caveat is documented in the aDDM tutorial §7 + the plot_model_cartoon docstring; this issue tracks the real fix.

The three decouplings

1. Regenerated predictive uses the default continuation. plot_model_cartoon deep-copies idata, collapses the posterior to its per-trial mean, and _make_idata_mean_posterior deletes the posterior_predictive group (src/hssm/plotting/model_cartoon.py:351-352). _use_traces_or_sample (src/hssm/plotting/utils.py:511-526) then regenerates it via model.sample_posterior_predictive(idata=..., data=..., inplace=True, draws=n_samples) with no continuation_mode/continuation_params → always the config default prolong_last_fixation. A user's per-call sample_continuation is discarded (it lives only in the user's original idata, which the cartoon deep-copies and never reads).

2. The drawn histogram + trajectories self-sample fixations (Mode 1). plot_func_model re-simulates via simulator(model="addm", theta=θ̄, ...) (model_cartoon.py:901, 938, 968) with no extra_fields, so the aDDM simulator self-samples its own gaze sequence instead of conditioning on each trial's observed fixations. The trajectories are genuine aDDM drift paths (metadata["trajectory"]), but for invented fixations, not the empirical ones. extra_fields never appears anywhere in model_cartoon.py.

3. No public API to pass either. plot_model_cartoon exposes neither extra_fields nor continuation_mode/continuation_params.

Net: the aDDM cartoon's predictive is neither the observed-fixation-conditioned PPC (model.sample_posterior_predictive, Mode 2) nor steerable by continuation policy — it silently differs from tutorial §6.

Proposed fix

• Add continuation_mode / continuation_params params to plot_model_cartoon → plot_func_model, and pass them to the simulator(...) calls (model_cartoon.py:901/938/968) and the step-1 regeneration.
• For covariate models (aDDM), thread the observed extra_fields (per-trial covariate rows aligned to the sampled θ rows) into those simulator(...) calls (Mode 2); tile the observed fixation set to n_samples to match the θ tiling.
• Test: assert the cartoon's predictive differs between prolong_last_fixation and sample_continuation, and that swapping observed r1/r2 changes the drift (stimulus-conditioning holds).

Acceptance

• plot_model_cartoon(model, idata, continuation_mode="sample_continuation", ...) yields a cartoon whose predictive matches a Mode-2 sample_continuation PPC.
• The cartoon conditions on the observed fixations by default (Mode 2), matching aDDM tutorial §6.