Added allow_gc Theano flags in the list of Theano flags.

Fixed crash in constant folding optimization

This is because either:
- There will never be any client and thus the node will be pruned, so
  performing constant folding would be a waste of time, or
- Someone will add clients later and will re-run the constant folding
  phase, in which case we will want to use the client-dependent logic to
  decide whether constant folding should be applied or not.

Fix tests

Fix import issue Olivier caught after my last PR was merged

Fix for a source of non-determinism

Corrected a few typos in scan documentation.

Fix optimization warning due as we didn't returned the right dtype for e...

This can just cause confusion.

Ready to merge: Determinism fix

str(type(foo))

Numpy's eigh(a, 'L') (eigh(a, 'U')) is a function of tril(a) (triu(a))
only.  This means that partial derivative of eigh(a, 'L') (eigh(a,
'U')) with respect to a[i,j] is zero for i < j (i > j).  At the same
time, non-zero components of the gradient must account for the fact
that variation of the opposite triangle contributes to variation of
two elements of Hermitian (symmetric) matrix.