Lsqfit

========================
Improvements to lsqfit.MultiFitter.

- New wavg that can average arrays of different shape (but same dimension).

- New option (wavg_all) for chained fits to average results from all the
fits at the end, to restore (some) correlations, when fast=True.

- New fit.format() for chained fits that includes summary information
about individual fits in the chain.

=======================
The extend parameter removed from nonlinear_fit; the associated
functionality (non-Gaussian priors) is now available automatically. The new
implementation, which resides entirely in gvar.BufferDict, is much simpler
and more robust. This causes changes in a variety of pieces of code, but will
have little or no effect on code that didn't explicitly use extend (and that
code should just stop referencing the keyword).

=========================
Another minor bug fixed (in wavg) that can cause problems in MultiFitter.

=========================
Two minor bug fixes in MultiFitter.

=======================
New implementation of lsqfit.MultiFitter that offers new features but also
is (slightly) incompatible with previous versions. The most important
new feature is the ability to change fitter keyword values in the middle
of fit chain, by including a dictionary with the new values as
an entry in the models list. This allows far more control over the
individual fits in chain of fits.

The incompatibilities mostly affect chained fits and are mostly internal. The
list of chained fits (fit.chained_fit) now contains all fits, including the
wavg fits done after a parallel fit. The output from fit.formatall() differs
accordingly. Another change for all fits: bootstraps are initiated using the
fit (not the fitter, as before) -- eg, fit.bootstrapped_fit_iter(10).

Other changes:

- maxit=0 normally results in one iteration of the nonlinear_fit fitter.
As per a request from Jason Chang, this is no longer the case when linear
terms are specified (using the linear keyword). Then the prior is used for
the nonlinear terms, and the linear terms are determined using the linear
least-squares formula (variable projection). This is an efficient way to
determine linear parameters when you have good estimates (eg, from previous
fits in a chain) for the nonlinear parameters.

- Use of a prior in lsqfit.wavg is deprecated. A prior never made much sense.

======================
- lsqfit.nonlinear_fit has a new keyword: linear. This can be used
to specify a list of fit variables that appear linearly in the fit
function. These are removed from the fit using variable projection, thereby
reducing the number of fit parameters; they are reintroduced after the fit.
This parameter is optional and currently somewhat experimental.

- nonlinear_fit methods load_parameters and dump_p are deprecated; use
gvar.load and gvar.dump instead. Similarly dump_pmean is
deprecated.