Instead of using current size of the graph, use the size the graph had
at its largest. This enables optimizations like constant-folding to
reduce greatly the total number of nodes in the graph without triggering
an error.

Mixed changes

Test nvidia driver

This allow to make sure the test for the driver was executed.

raise an exception if the test fail. This is because some nvidia driver version don't correctly handle our reduction code.

Do not repeat a variable in a function's inputs

Simplified code for In.borrow and Param.borrow

That's the reason why DebugMode was complaining.

Broadcast input of sparse_from_dense so ndim=2

Fix dtype of input value in test_diag

Since csc_matrix's default dtype is floatX, the numpy value needs
to have a compatible dtype.

Do not constant-fold Alloc if it is the output

Since mutable=True is incompatible with borrow=False, it is better to
raise an exception to make sure it's being noticed.

make_node used to crash, now it works.
A test detects the problem.

This makes things easier for shape inference and gradient operations.
Also fix broadcasting pattern of the gradient.

This commit simplifies the code so that In.borrow and Param.borrow can
only be True or False (not None). It also makes the corresponding code
for these two objects more similar.

There should be no functional change, it just makes the code easier to
understand.

Finished DOC/TEST/REVIEW sparse Diag Op

Fix test_diag_grad to work with verify_grad

Also fix an optimization bug in that case.

Fix perform of structureddot_grad with dense grad