This commit fixes the chi2 analysis so that it is no longer biased.
Previously, the chi2 analysis pull plots showed a consistent bias. At
first, I thought this was due to the fact that the posterior wasn't
gaussian, but even after switching to percentile plots based on the
algorithm outlined in "Validating Bayesian Inference Algorithms with
Simulation-Based Calibration", I was still seeing a bias. I finally
tracked it down to the fact that I was applying the energy scale
parameters to the data instead of the Monte Carlo. Therefore, in this
commit I update the posterior to now apply the energy scale parameters
to the Monte Carlo instead of the data. This has the slight disadvantage
that the final histograms will be binned in the biased energy, but
that's not really a big deal.

In addition, this commit contains several other updates:

- switch to plotting percentile plots based on the algorithm in
  "Validating Bayesian Inference Algorithms with Simulation-Based
  Calibration" instead of pull plots
- apply both the energy scale and resolution at the individual particle
  level, i.e. there is no longer an energy resolution term for electron
  + muon fits
- separate pull plots and coverage plots. Previously I was making both
  the p-value coverage plots and the pull plots at the same time.
  However, the pull plots shouldn't have anything to do with the GENIE
  weights whereas the p-value coverage plots should draw samples
  weighted by the GENIE weights. In addition, for the pull plots we draw
  new truth parameters on every iteration whereas for the p-value
  coverage plots we only draw them once.
- switch to using KDEMove() for the MCMC since I think it samples
  multimodal distributions a lot better than the default emcee move.
- I now correct for the reconstruction energy bias in plot-michel and
  plot-muons

This commit updates how the energy bias is applied when we correct for
the energy bias in correct_energy_bias(). The correct way to apply this
correction is to compute:

    T_corrected = T_reco/(1+bias)

whereas previously we were multiplying by (1-bias).