TODO

Current bugs/limitations:
-------------------------
- scanning over the frequency variable on the CUDA device is impossible

TODO
----
- make it possible to run more than 64k * 64 threads at the same time
- automatically finish run when the result is not changing?
- doc: compare efficiency of the different RNGs
- why adding more precision to the factor results in nans?
  - add support for RNG testing mode.
- add support for double precision calculations
- the folding of arguments of trigonometric functions should be to -pi, pi to make the intrinsic functions more precise
- the folding should be possible to do directly in the kernels
- optimize state dumping for large scans.
- restore proper histogram in bistable via masked arrays
- make prev_state_results a named tuple
  
RNG test results:

# time ./bistable.py --force=0.5 --gamma_=0.5 --psd=1.0 --d0=0.025 --spp=1000 --paths=8192 --simperiods=10000 --samples=100000 --rng=kiss32

xs32:   0m11.271s
nr32: 	0m29.358s
kiss32: 0m23.508s


# time ./bistable.py --force=0.5:0.7:50 --gamma_=0.3 --psd=1.0 --spp=200 --rng=nr32 --paths=16384 --simperiods=1000 --samples=20000 --output=/tmp/g0.3_d${NOISE}_kiss32 --output_mode=path --d0=${NOISE} 
kiss32: real	0m45.177s
nr32: real	0m56.128s