TST: accept small error in threaded random test

freebsd and windows change x87 precision mode (fctrl bit 8 and 9) from
extended to double in child threads so the random numbers cannot be
exactly the same from master and child threads.
see gh-4909

rho results in 0. for kappa < 1.4e-8 whch leads to nans appearing and an
infinite loop.
the second order taylor expansion is more precise up to at least 1e-5.
Closes gh-4720

mtrand accepts seeds larger than 32 bit but silently truncates them back
to 32 bit. This can lead to accidentally getting the same random stream
for two different seeds, e.g. 1 and 1 + 2**40.

The random module currently relies on the GIL for the state
synchronization which hampers threading performance.
Instead add a lock to the RandomState object and take it for all
operations calling into randomkit while releasing the GIL.
This allows parallizing random number generation using multiple states
or asynchronous generation in a worker thread.

Note that with a large number of threads the standard mersenne twister
used may exhibit overlap if the number of parallel streams is large
compared to the size of the state space, though due to the limited
scalability of Python in regards to threads this is likely not a big
issue.

Adds check with np.isnan(p) and raises ValueError if check is positive.

Explicitly Test that the default shape does not raise a
DeprecationWarning.

Check that a covariance matrix that is not positive-semidefinite
raises a RuntimeWarning.

Closes gh-3173

closes gh-4270 and gh-3263
also regenerate with cython 0.20.1