Scalarize input

[gustaphe]

This is a breaking change, but one that I really feel makes LsqFit more Julian. Would close #145

Instead of

model(x, p) = exp.( - ( x .- p[1] ).^2 ./ p[2]^2 )

the model definition (the operation which involves the user the most) becomes

model(x, p) = exp( - ( x - p[1] )^2 / p[2]^2 )

i.e. the user doesn't need to think about the shape of x, they just define the behavior with respect to a scalar.

From the tests, it looks like you are just moving the complexity from the model definition to the data generation
(which goes from y = model(x, p) to y = model.(x, Ref(p))), but note that in real world cases you wouldn't be generating the data from your own model like this.

One part that I'm a bit uncertain about is multivariate models: I've currently written it so that the x argument for curve_fit is a AbstractArray{AbstractArray} (e.g. [ [1.0, 3.0], [2.0, 3.4], [1.5, 2.0] ] for a two-variable function with three datapoints). One might want to go through and change this to accept matrices, I just haven't (yet). There are afaict currently no tests for multivariate fitting either.

Oh, and the tests weren't passing (some of them even crashed) before, so I relaxed them a bit in the first commit.

Are you interested?

[pkofod]

Yes. You are right that it's a breaking change, so it would have to be postponed a bit, but I think it's a better interface.