interval CDF differences for interval censoring/truncation

[bob-carpenter]

It'd be useful for both efficiency and ease of notation if we had a function defined by:

foo_cdf_diff_log(U,L,theta) 

  =def=  log(foo_cdf(U,theta) - foo_cdf(L,theta)) 

  =  log_diff_exp(foo_cdf_log(U,theta), foo_cdf_log(L,theta))

The negation of this is what gets added when T[L,U] is added to probability function foo.

Then we'd use those for censoring and truncation instead of the expression the right, which we're using now. It should be done with a log-sum-exp type operation. Even a utility implementation that just used Stan's existing log_diff_exp function would be useful.

[syclik]

We should wait on stan-dev/math#320 before implementing this.

[syclik]

We'll also need to add functions in math.