Allow lambda zero for parametrization in Exp distribution.

[saona-raimundo]

This allows the parameter lambda = 0 in the distribution Exp.

Implementation

Taking advantage that 1 / 0 = infinity, we change nothing in the implementation, i.e. a sample from Exp(lambda) is the result of a sample from Exp1 multiplied by 1 / lambda.

Performance

This makes the sampling from Exp(lambda = 0) lose time sampling from Exp1 when there is no need since the result is always infinity. This is okay since it is not the common case.

Documentation

The error LambdaTooSmall accounts now only for negative or NaN parameters lambda.
Also, In the documentation for Exp, there are two changes: the explanation of density and constructor new. Please check the style of this.

In the explanation of the density, we included the case lambda = 0, assuming that the user is using primitives (most of the cases).

In the constructor new, a remark was added. It seems a bit overly complicated, but this is because of the trait N: Float. Since Float has no infinity method, we rely on the behaviour of 1 / 0 of the custom type N. I made this explicit by explaining the implementation itself! I did not come up with a better explanation.

[dhardy]

Looks good other than that little nit.

[dhardy]

Meanwhile: statrs-dev/statrs#102

This particular degenerate case (exponential with rate 0) is a clear extension of the regular model.

[vks]

Looks good!

[dhardy]

Well, lets merge this at least: this particular degenerate case is not problematic (provided one accepts that one cannot necessarily integrate over an infinite domain and get the expected result).