dtype for different Tensor CAReduce operations

Ops:GetItem2D&GetItemScalar to return a subtensor/scalar from a sparse m...

gpu accelerated tensordot by conversion to matrix product

Ops:GetItem2D&GetItemScalar to return a subtensor/scalar from a sparse matrix tested. grad() not implemented because too complicated for now

Allow data for sparse matrices to be broadcastable

A problem occurred when the data (and indices) were constant vectors of
length one, since they had a broadcastable pattern of (True,), not
(False,) as was expected. That happened in the buildbot in debug mode
with seed 24856.

Tests for tile (and some fixes besides)

TileGrad implements only one very narrow special case, and it implements
it wrongly. Raise NotImplementedError instead of using this botched
implementation.

Raise a proper error when reps is not a constant (since a tensor value
is not supported by this implementation, as it is) and also force
x.ndim == len(reps), since bad things happen if it's not the case.

Hide warnings for a past bug in the buildbot

Fix subtensor crash

Fix runtime crash with the local_subtensor_of_alloc optimization when the input have a broadcastable dimensions.

Edit NEWS.txt

Doc op

Also add ".prod()" method to tensor variables.

This enables specifying the dtype in which the internal summation
should be done, like in Numpy.

Also add a tests for one specific case.
The usual cases should be tested by the regular tests, since the sum
nodes' dtype are no longer allowed to be None.

Update existing test to check that a TypeError is raised if an
inappropriate output dtype is specified in sum().

It's easier to understand the doc if we have explicit kw args.

If the output dtype would force a downcasting of the values
(for instance, summing a float vector into an integer), a TypeError
is now raised.

Also, a Sum node has to have a non-None dtype. It is checked in
make_node because for some reason the pretty-printing mechanism needs to
instanciate "Sum()".