Compiler

Better error message if problem in givens

Fix max'd error in EquilibriumOptimizer with constant folding

Now constant folding does not bust it anymore.

It used to wait for the first attempt to apply an optimization
after it has already crossed the threshold.
Now, it signals that the optimizer should stop earlier, just after
crossing that threshold.

This is needed as we changed the signature of some function and thus we should force the recompilation of all gpu op.

It seam more logical to associate it with the compiler then the Type. This also make cuda_ndarray.cu compiled with it.

make {nvcc,gcc}_module_compile_str a class with another function compile_args that get added in the keys.

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.

Fixed circular import issue

Sorry, I should have tested more thoroughly this stuff with GPU before
pushing it :/

Fix tolerance in Usmm test.

Added missing import

Fix when forcing cuda to be disabled

Fixes for Python 2.4

The GpuOp class must be available even when Cuda is disabled (e.g. by
forcing cuda_available = False). Otherwise some tests fail.

This commit moves the definition of GpuOp directly into cuda/__init__.py
so that it works.

Minor fixes

Update NEWS.txt

Make blas functions work with all stride patterns

Fix dtype in sparse test.

Mixed changes