Debug and fix the random number generator degeneracy problem

[fangq]

Users have reported occasions where MCX produced a streak of high-intensity pixels along the forward scattering direction, likely caused by the failure of the built-in LL5 RNG.

If Logistic-lattice based RNG is not stable for productivity purposes, need to consider other alternative RNGs, such as cuRAND

https://developer.nvidia.com/cuRAND

or revive the MT RNG included in MCX

[fangq]

Update: the forward-scattering streak may be a result of incorrect floating-point epsilon setting.

The IEEE 754 epsilon for float32 is around 1.1e-7. Setting it to 1e-10 will likely drive the code into the wrong if-block as shown here

https://github.com/fangq/mcx/blob/master/src/mcx_core.cu#L394

the small epsilon issue was recently corrected in

http://sourceforge.net/p/mcx/svn/341/

hope this patch will also make the forward-scattering streak go away, but because it has not been consistently reproducible, rigorous test has not been done yet.

nonetheless, exploring other RNGs options are still needed.

[fangq]

Two new RNGs are now implemented in MCX, including POSIX erand48 and xorshift128+.

b14ccf1

while we are continue investigating the quality of LL5 RNG, the priority of this task is reduced.

close this for now.