Fix sparse scaling matrix in quadpotential

[aseyboldt]

Non-diagonal scaling matrices seem to be broken at the moment, as they do not work with theano. This pull request should fix that.

Also, think the sparse implementation was incorrect. velocity and random did not match each other.

[twiecki]

Thanks @aseyboldt! Looks like some tests failing under python 2.7:

======================================================================
FAIL: pymc3.tests.test_quadpotential.test_elemwise_velocity
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/travis/miniconda2/envs/testenv/lib/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/home/travis/build/pymc-devs/pymc3/pymc3/tests/test_quadpotential.py", line 48, in test_elemwise_velocity
    assert np.allclose(v(x_), 1/scaling)
AssertionError
======================================================================
FAIL: pymc3.tests.test_quadpotential.test_elemwise_energy
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/travis/miniconda2/envs/testenv/lib/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/home/travis/build/pymc-devs/pymc3/pymc3/tests/test_quadpotential.py", line 61, in test_elemwise_energy
    assert np.allclose(energy(x_), 0.5 * (1/scaling).sum())
AssertionError

[aseyboldt]

Ah, interger division... should work now.

[twiecki]

Thanks for the fix and unittests!