Support non-filtered `predict_times` for trajectory-based evaluation

[mattlevine22]

• To start, I suggest we support the case where all forecast_times > all obs_times
• do sample(..., DynamicalModel, obs_times, obs_values, forecast_times)
• for now, interpret using the inference method (filters can do the forecasting, and unrollers can keep unrolling)
• CD_Dynamax cdnlgssm and cdlgssm (not cdnlssm) offer filter_and_forecast

[DanWaxman]

Here are decisions we made of expected behavior under three cases:

Cases

Case 1

with Simulator(...):
    model()

Case 2

with Simulator(...):
    with Filter(...):
        model()

Case 3

with Filter(...):
    model()

We agreed that Filter outside of Simulator was nonsensical.

Expected Behavior

These are partially dependent on the new choice of #129 to have a two-layer sample call.

Input Provided	Case 1: Simulator → Sample	Case 2: Simulator → Filter → Sample	Case 3: Filter → Sample
observed_times, observed_values, predicted_times	diffrax solve on observed_times ∪ predicted_times; observations = None for predicted_times! (Note: Easier if no inference via SDE Simulator)	Filter consumes observed_times and observed_values. Passes filtered_times and filtered_probability_distributions forward to simulator. Solve Chapman-Kolmogorov equations for filtered_times → predicted_times.	sample_intp returns all unconsumed times as well; sample checks that all times are consumed!
observed_times, observed_values	Same as current, except for SDESimulator which should error.	Filter consumes observed_times and observed_values; Simulator no-ops.	Same as now.
observed_times, predicted_times	Error! (observed_times always needs observed_values)	Error! (observed_times always needs observed_values)	Error! (observed_times always needs observed_values)
predicted_times (Current behavior of observed_times w/ no observed_values)	Fine	No-op in Filter → falls back to Case 1.	No-op
observed_times	Error! Always generate with predicted_times now.	Error! Always generate with predicted_times now.	Error! Always generate with predicted_times now.