exploit new expint(n,z) function

[stevengj]

Might be good to define a few additional functions in terms of the new expint functions:

• the Ei function (probably call it expinti) (add expinti function #257)
• the sinint and cosint functions for complex arguments
• incomplete gamma functions for complex arguments: upper incomplete gamma function Γ(a,z) #271
• the logarithmic integral function li(x) = Ei(log(x)) (too trivial? or does it do something interesting for complex args?).
• expintx(ν,z) = exp(z) * expint(ν, z): expintx function — exponentially scaled expint #249
• log of the incomplete gamma function log Γ(a,x), using expintx as commented here: WIP: add loggamma(a,z) #283

(Even for real arguments, the expint implementation seems to be faster than what we have now for gamma_inc, although it might not be as accurate for the "lower" incomplete gamma function. I feel like we should just have a gamma(n,x) method that returns the upper incomplete gamma function, similar to Mathematica, in any case.)

[azev77]

Thanks @stevengj !
This makes it possible to add the requested Gompertz distribution to Distributions.jl.

1. The moment generating function for Gompertz requires:
   
   Is this available here?
2. Currently expint() doesn't work (unless I'm using #Master) bc the latest release was in May.
   Approx when can we expect a new release?

[stevengj]

1. Yes, expint(t/b, η) should work fine for that case. The new implementation by @augustt198 supports arbitrary real and complex orders and arguments.

2. I think we need to fix expint type instabilty for negative order #246, hopefully over the next week or so, before a new release.

[stevengj]

It looks like what you really need for the Gomperz CDF is the scaled function expintx(ν, z) = exp(z) * expint(ν, z). It could be useful to pull this out as a specialized function, since the expint implementation internally scales by exp(-z) in many cases — it is better to avoid computing exp(z) * exp(-z), in order to avoid spurious underflow for large arguments. We could probably refactor expint to pull out the unscaled code.