Make NumpyOps CPU kernels generic (#627)

* Make NumpyOps CPU kernels generic

This PR makes most CPU kernels generic, so that they can take both
float32 and float64 arrays (and hopefully in the future float16). I
experimented with kernels in Cython + fused types and kernels as C++
with templates, I found the C++ template route more promising:

- More compact/ergonomic implementations with fewer compile-time
  conditionals.
- Opens up the possibility to easily use SIMD intrinsics in the
  future.

To allow genericity in the NumpyOps method arguments, we use:

- Fused types when we require a specific dimensionality;
- np.ndarray otherwise.

Some of the kernels are not made generic:

- cpu_scatter_add: needs tests to verify that the op still works
  correctly.
- cpu_position_encode: the position_encode op doesn't take float
  array(s).
- lstm kernels: I need to look more deeply into them.

* Include C++ headers in sdist

* NumpyOps: Use workaround for cython/cython#4697

* Namespace-qualify memcpy

* ReLU kernel: never output -0.0

* Add fixes suggested by @svlandeg

Author: danieldk

MANIFEST.in

@@ -1,4 +1,4 @@
-recursive-include thinc *.cu *.pyx *.pxd
+recursive-include thinc *.cu *.pyx *.pxd *.hh
 include LICENSE
 include README.md
 prune tmp/

setup.py

@@ -113,7 +113,7 @@ def setup_package():
         version=about["__version__"],
         ext_modules=ext_modules,
         cmdclass={"build_ext": build_ext_subclass},
-        package_data={"": ["*.pyx", "*.pxd", "*.pxi", "*.cu"]},
+        package_data={"": ["*.pyx", "*.pxd", "*.pxi", "*.cu", "*.hh"]},
     )