mir.random.discrete - fix for zero probilities

[wilzbach]

(in reference to #235)

As @WebDrake pointed out Discrete doesn't handle zero probabilities well.
Also as we don't use chi-squared tests, but I think it's at least helpful to insert the assert's for the currently tested distributions.

With zero probs we can't use lowerBound, because it will it will not stop on the first element, but go to all equal element. Hence we either count the first category twice or the third one.
Here's an example of what's going on:

import std.stdio;
import std.range;
auto cdPoints = [1, 3, 3, 4];
foreach (el; 0..5)
{
    writeln(el, "->", cdPoints.assumeSorted!"a < b".lowerBound(el));
}

foreach (el; 0..5)
    writeln(el, "->", cdPoints.assumeSorted!"a <= b".lowerBound(el));

0->[]
1->[] // would count first category twice
2->[1]
3->[1]
4->[1, 3, 3] // only for "[]"

0->[]
1->[1]
2->[1]
3->[1, 3, 3] // would count third category (the zero one)
4->[1, 3, 3, 4] //only for "[]"

[wilzbach]

the split looks too random with 1000

[codecov-io]

Current coverage is 96.53%

Merging #236 into master will increase coverage by <.01%

@@             master       #236   diff @@
==========================================
  Files            19         19          
  Lines          3253       3262     +9   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
+ Hits           3140       3149     +9   
  Misses          113        113          
  Partials          0          0          

Powered by Codecov. Last updated by 220b2c8...4a03dd9

[9il]

OK, but don not use Random alias then. Use exact name instead.

[WebDrake]

Agree, it's better to specify an exact generator just in case someone later adds some tests that depend on the pseudo-random sequence.

[wilzbach]

OK, but don not use Random alias then. Use exact name instead.

Ok - done ;-)