Flavio

This is a quick bug fix release necessary due to an incompatible change in numpy v1.12 [released](https://github.com/numpy/numpy/releases/tag/v1.12.0) two weeks ago that made processes depending on B→P form factors stop working (thanks to Stefanie Reichert for reporting).

This first release in 2017 brings lots of new features and some bug fixes.

Several improvements have been made to probability distributions, which are required to quantify both theoretical and experimental uncertainties needed to compute prediction uncertainties and to construct likelihoods for fits:
- A new class [GammaUpperLimit](https://flav-io.github.io/apidoc/flavio/statistics/probability.m.html#flavio.statistics.probability.GammaUpperLimit) appropriate for low-statistics measurements resulting in an upper limit (e.g. on a rare decay branching ratio),
- A new class [MultivariateNumericalDistribution](https://flav-io.github.io/apidoc/flavio/statistics/probability.m.html#flavio.statistics.probability.MultivariateNumericalDistribution) that allows to specify _any_ probability distribution in an arbitrary number of dimensions. This works with all the functions implemented, like computing uncertainties and Bayesian fits.
- Constraints in the form of a fixed range (uniform likelihood within and zero outside) can now be specified in YAML files in the intuitive form `[0.1, 3.5]`.
- A new way of specifying arbitrary probability distributions in YAML files.

There is also a new page in the documentation about [details on the treatment of probability distributions](https://flav-io.github.io/docs/probability.html) and how to define them in YAML. The [API docs](https://flav-io.github.io/apidoc/flavio/statistics/probability.m.html) have been updated as well.

On the physics side, the following processes have been added:
- The forward-backward asymmetry in inclusive B→X_ql⁺l^- decays. Like The branching ratio added in [v0.15](https://github.com/flav-io/flavio/releases/tag/v0.15), this is based on [arXiv:1503.04849](https://arxiv.org/abs/1503.04849) and profited a lot from help by Tobias Huber.
- The branching ratio of B_c→lν.
- Lepton flavour ratios in B→Mlν decays, including experimental measurements of the μ/τ ratios in B→D^(*)lν.

Finally, the following bugs have been fixed:
- In the presence of new physics, the decay B_d→l⁺l^- erroneously used b→s Wilson coefficients. Thanks to jackypheno (Jacky Kumar) for spotting and fixing this.
- Measurements with a Gaussian upper limit constraint returned the wrong likelihood if the confidence level specified was different from 68%.
- A glitch in the log-enhanced QED corrections to B→X_ql⁺l^- was fixed (the change is numerically irrelevant).

The main new feature in this release is the addition of the branching ratio of the inclusive semi-leptonic B decay B→X_clν, with l=e or μ (or their average). Many thanks to Paolo Gambino for useful discussions and for sharing numerical results necessary for implementing and validating this! See [arXiv:1411.6560](https://arxiv.org/abs/1411.6560) and references therein for details on the theory.

An additional, minor improvement since v0.16.1 concerns the treatment of correlated ucertainties in Bayesian fits; see commit a32c80d45e14522dd5c2e11d942512ad9922dfe7 for details.

Finally, the default values of masses and widths have been updated from PDG 2015 to PDG 2016 values. There has not been any numerically relevant change for the currently implemented observables.

This bug fix release fixes
- a regression introduced by the recent internal change in the `Constraints` class in v0.16 that broke the fit functions,
- an inconsistency in the combination of experimental likelihoods in the presence of multiple experiments measuring the same observable when using the `FastFit` class.

Additional unit tests have been added to improve the stability of these fitting functions.

This release focuses on stability improvements with only few new features.

Improvements on the physics side:
- Added lepton flavour ratios for B->Plnu decays, e.g. `<Rtaumu>(B0->Dlnu)`
- Added lepton flavour ratios for B->Xll decays, e.g. `<Remu>(B->Xsll)`
- Added measurements of B->K*mumu "optimised" angular observables (P') by LHCb

Other improvements:
- More consistent behaviour of `flavio.plots.band_plot` when using the `pre_calculated_z` and `interpolation_factor` options (see [API docs](https://flav-io.github.io/apidoc/flavio/plots/plotfunctions.m.html))
- `flavio.plots.q2_plot_exp` now properly handles asymmetric experimental uncertainties

In addition there have been a few under-the-hood changes in preparation for future features.

This minor release fixes the following bugs:
- An erroneous factor of 1/2 in the angular observables A_FB, F_H, as well as the direct CP asymmetry in B->Pll decays was removed. Note that these quantities are compatible with zero in the SM, so this is relevant in the presence of new physics. Thanks to Christoph Niehoff for pointing this out.
- The plot function `q2_plot_th_bin` now also works for non-SM Wilson coefficients (24, thanks f-thiele)

As a new feature, a function `np_uncertainty` was added that works just like `sm_uncertainty` but allows to specify non-SM Wilson coefficients. This allows to compute the uncertainties in the presence of new physics (see [API docs](https://flav-io.github.io/apidoc/flavio/functions.m.html#flavio.functions.np_uncertainty)).