lgamma slow with defaults

The `boost::math::lgamma` implementation is about 2-3x slower than the `std::lgamma` implementation when used with standard `double`s.

When turning off the do the propagation from `double` to `long double` the boost implementation is comparable in speed to `std::lgamma`, but still 10-20% slower.

However, when turning off propagation to `long double` problems with early overflow are the price to pay... see the next issue.