To be able to use negative offsets in indexing, threadIdx.x (and y)
should be cast to signed int.

It may be slower than allocated new memory and calling memset,
but if that non-contiguous memory was provided, it is probable
that the output will end up being copied into that non-contiguous
memory, and that would be even slower.

In DebugMode, do not use the no_recycling list when calling the CLinker
to build individual thunks, so that output memory can always be re-used,
even for outputs of the actual function.

This enables check_preallocated_output to work on Ops that compute
the function's outputs.

This should close gh-1174.

Speed up the gradient of AdvancedSubtensor1 WIP

Add information to error message.

Fix infinite opt loop

Lazy import of sparse module and rename sparsegrad according to python coding style.

Even in DebugMode, even before the inplace phase.

It helps locating infinite loops

Add a flag to decide which gradient should be returned from AdvSubtensor1

Tensor Type has now a sparse grad attribute which can be set or unset to determine if sparse gradient is to be enabled or not. Note if needed, we can provide getters and setters to set the attributes to provide more clean interfaces to application. Right now application can access this attribute and set it before calculating gradient

Ready for merge: Streamline dot & tensordot / Allow dot products of n-dimensional variables

Fix bugs reported by the DL tutorial wrote by Nicolas

Fixed a race condition.

Importing Scipy sparse in the ContructSparse Op to avoid unnecessary Scipy dependency.

Fix crash tensor.roll(Var, iscalar) reported by Jeremiah Lowin.

fix gh-1154