No warning if g++ does not support -march=native

Instead of detecting llvm, we try to use -march=native, which is more
general.

Fix C code of GpuDownsampleFactorMaxGrad.

Use safe_new(x) instead of x.type() in scan grad.

Simpler graphs for canonical slices

This copies the test_value, making possible to use a test value
for sequences. Add test.

Compute hash on whole source file

Fix insertion of -arch=... when compiling with nvcc

Minor fixes

Fixes gh-1306

It used to be inserted only for the first Op after GPU initialization.

It was causing errors when running with cuda-memcheck, or when running
on Mac OS X.

including support code, in dynamic modules.

* When `step` is 1 and either `start` or `stop` is a boundary, the
  canonical representation of `slice(start, stop, step)` is now simpler

* Other simplifications in the more general case when `step` is a
  constant

* Fixed crash when calling `get_scalar_constant_value()` on a numpy
  array.

* Ensure that start <= stop in canonical slice representations, which
  makes it easier to compute their length.

Fix ConstructSparseFromList.infer_shape

Now, the first argument contains the shape itself, so we should use
its value, not its shape.

Fixes

Add assert in make_node

fix bug reported by Nicolas

Fix bug in erfinv GPU optimization

Check Python long are OK iff they fit in int64