Lsqfit

==========================
- Fixes installation issue with Python 3.10.

- Documentation fixes (thanks to Nolan Miller and Evan Berkowitz).

=========================
Bug fix for lsqfit.gammaQ when gsl is not installed.

Version 12 2021-08-14
======================
- Removes BayesIntegrator (and BayesPDF). It is replaced by
the more capable PDFIntegrator in the vegas Python module.

- Adds two new methods to nonlinear_fit: logpdf(p) and pdf(p).

- Fixed minor bug when GSL not installed and maxit=0.

- Fixed bug involving noise parameter when no prior used.

=======================
Small update to fix a bug and add a feature. Thanks to Nolan Miller
for both.

- Fixes bug in empbayes_fit(z0, ...) when z0 is or contains integers
instead of floats.

- MultiFitter.lsqfit has a new keyword, chained. If chained=True
chained_lsqfit is used in place of lsqfit.

- MultiFitter objects have a new method MultiFitter.empbayes_fit that
is analogous to lsqfit.empbayes but for MultiFitters.

=========================
Minor update responding to a feature request.

- fit.evalchi2(p) evaluates the chi**2 for the fit at an
arbitrary value p for the fit parameters.

- Bug fix so maxit=0 works again.

========================
- Additional sanity checks added for debug=True in nonlinear_fit.

- nonlinear_fit now uses gvar.regulate rather than gvar.svd to
deal with singular correlation matrices and to decompose the
correlation matrices when forming the chi**2 function.
To regulate the correlation matrix, one now specifies
either eps or svdcut. The eps regulator uses Cholesky
decomposition, while the svdcut uses SVD decomposition
(the only option previously). Cholesky decomposition
is faster but less robust. Keywords add_svdnoise and
add_priornoise are replaced by addnoise, but the old
keywords continue to work (for legacy code).

- Tweak to reporting on stopping conditions.

- Updated documentation.

- Requires gvar v11.6 or later.

=========================
Minor fix for compatibility with gvar v11.5.2.