added python binding for cuda nd array dimshuffle
added test of the above
used the above to cover all four tranpose cases of test_cuda_ndarray.py:test_dot

fixed an issue where test_reshape in test_cuda_ndarray did not use the correct seed for random number generation

These tests were previously computing wrong values.

condition.

one that works on scalars.

previously introduced function for merging two slices.

Test that gemv are inserted into the graph. Disable the gemv test on complex as they don't get inserted.

re-use it for other optimizations.

optmized right now, and so they make no sense .. The code asks now for more
tests though.

interesting ones ?) with the disadvantage of making it slow

combinations of two subtensors.

include the length of the tensor as possible ``stop`` for a useless
subtensor.

changes done to the canonical form of the slice.

slices given the idx_list and the list of node inputs. The other function
transforms a slice into a canonical form ( where all entries are positives),
which makes dealing with slices much easier. I've also refractored the
perform of the Subtensor to use the first function ( since is mostly copy
paste from that code anyway)

use the gcc bug workaround in all case to make it more consistent and fix a bad code generation case.

This does not change the logic of the Gemm optimizer, so #415 might
still be needed in the future.

This change:
  - apply the following optimization whether or not the transpose is inplace:
T.dot(X,Y).T -> T.dot(Y.T, X.T). The latter form is expected by the Gemm
optimizer.
  - removes the Warning class raised in test_blas.py:just_gemm, because
it was caught silently, hiding the fact that another case was failing.
A "Failure" is now thrown instead;
  - adds another test case for transposition;
  - makes the first case in test_blas.py:test_inplace0 expect a
"gemm_no_inplace" instead of _dot22.