Implement poisson function

[denisalevi]

This PR implements the poisson function. Should close #174.

See brian-team/brian2#1111

[denisalevi]

There are two different ways that poisson random numbers are generated in CUDA now:

1. Using a buffer system which is refilled from host code in regular intervals using the cuRAND host API. This is used for poisson(lambda) when lambda is a the same value across all neurons/synapses and doesn't change during a run.
2. Using on the fly RNG from device code using the cuRAND device API. This is used for poisson(lambda) when lambda is a vectorized variable (different across neurons/synapses) or when it can change during a simulation. We do this because the host side API is only efficient when generating many random numbers for the same lambda.

To decide which RNG type to use, I regex match the <lambda> string in _poisson(<lambda>, _vectorization_idx) in the generated code. Stored as lambda_str = <lambda>, I first test if if is a literal and if not I look up the variable in the codeobjects variables dictionary and check if it is scalar. It looks something like this:

lamba_vectorized = True
try:
    # try to convert to float, will raise ValueError if not possible
    lamda_value = float(lamda_str)
    lamba_vectorized = False
except ValueError:
    # lamda is not a literal but a variable, check if lamda is scalar
    lamda_var = codeobj.variables[lamda_str]
    if lamda_var.scalar and lamda_var.constant:
        lamda_vectorized = False
        lamda_value = lamda_var.value

@mstimberg
Could you maybe review this check? For poisson(<lambda>) calls where <lambda> is not a literal, will it always be stored as a Variable in codeobj.variables? And if so: If I make check that it is scalar and constant, can I be sure that it won't change during a run and also that I always have access to its value before the standalone simulation is run?

If the check is fine, it will still miss a few cases where host side RNG could be used: E.g. if <lambda> is a non-scalar variable that happens to still have the same value for all neurons/synapses and doesn't change during a run. Such cases could probably be detected already on Python side (if a vectorized <lambda> variable is set with a fixed array or a scalar value) or on C++/CUDA side (if it is set using string syntax but happens to have same lambda for all neurons/synapses, similar to how we check if delays are scalar even though they were set with string syntax). Just leaving this here as a comment, not a priority for now.

[denisalevi]

I added poisson versions for all the seed RNG tests. The test_random_number_generation.py has a lot of tests now and the same test for each of the random number functions. I started refactoring it such that the tests are dynamically generated (same test but different RNG functions). Something wasn't working, I left the code as comment at the bottom of the file, for another time to figure out...

Tests are passing. I opened #228 to make sure the logic for checking poisson's lambda variable are correct.

Merging.