implement `sensitivity_analysis` helper method

implement a method which can easily perform sensitivity analysis on a DynODE model by selectively leaving out one parameter of interest out of the posterior samples dictionary at a time, and rerunning the posterior predictive distribution while re-sampling the parameter of interest.

basic sketch of the interface, subject to change.

def sensitivity_analysis(simulator : Callable, posteriors : dict[str, Array], parameters_to_vary : list[str]) -> dict[str : dict[str: Array]]:
"""
perform sensitivity analysis by calling `simulator` conditioned on `posteriors` while resampling each param within `parameters_to_vary`. 
return dict[str, dict[str, Array]] with the first key mapping to the varied parameter, and the second mapping to the posterior predictive with the resampled parameters.
"""
    for param_vary in parameters_to_vary:
           dict_leave_one_out= posteriors.copy()
           del dict_leave_one_out[param_vary]
           pred = Predictive(simulator, posteriors=dict_leave_one_out)
           ...
    pass

this method would basically just run a for loop (or parallelized pmap) over the parameters_to_vary and remove each from the posteriors dictionary before passing the one-removed dict along with simulator to the numpyro.Predictive class to regenerate more posterior predictive distributions with the left-out parameter being resampled.

Start with the simulator callable and the numpyro.Predictive usage described in the issue; no files or tests are named. Trace how posteriors and parameters_to_vary are represented, then verify that each requested parameter can be left out and resampled. Done means returning the specified nested mapping for every requested parameter.