Make MCMC calibration more modular for various calibration options

[btobers]

MCMC calibration is currently hard-coded to expect that the 0th element in the observation and model prediction lists is the glacierwide mass balance, and the 1st element is the 1d elevation change (if option_calib_elev_change_1d):

    def log_likelihood(self, m):
        log_likehood = 0
        for i, pred in enumerate(self.preds):
            # --- Check for invalid predictions early ---
            if torch.all(pred == float('-inf')):
                # Invalid model output -> assign -inf likelihood
                return torch.tensor([-float('inf')])

            if i == 0:
                # --- Base case: mass balance likelihood ---
                log_likehood += log_normal_density(
                    self.obs[i][0],  # observed values
                    mu=pred,  # predicted values
                    sigma=self.obs[i][1],  # observation uncertainty
                )

            elif i == 1 and len(m) > 3:
                # --- Extended case: apply density scaling to get binned elevation change ---
                # Create density field, separate values for ablation/accumulation zones
                rho = np.ones_like(self.bin_z)
                rho[self.abl_mask] = m[3]  # rhoabl
                rho[~self.abl_mask] = m[4]  # rhoacc
                rho = torch.tensor(rho)
                self.preds[i] = pred = (
                    self.preds[i] * (pygem_prms['constants']['density_ice'] / rho[:, np.newaxis])
                )  # scale prediction by model density values (convert from m ice to m thickness change considering modeled density)

                log_likehood += log_normal_density(
                    self.obs[i][0],  # observations
                    mu=pred,  # scaled predictions
                    sigma=self.obs[i][1],  # uncertainty
                )
        return log_likehood

This doesn't allow much flexibility in calibrating against additional datasets or turning on/off various calibrations. We should instead pass the observation and predictions as dictionaries to pygem.mcmc.mbPosterior with paired keys that get evaluated accordingly to make the sampling framework more robust.