Elemwise faster when c or f contig inputs

This speed up hard_sigmoid and ultra_fast_sigmoid by 2x.

We ignore input that are broadcasted scalar when checking for fortran
inputs.

Do not work yet when the inputs is fortran, as the outputs is always c contig for now.

Add hard_sigmoid, faster, but more approx.

Add ultra_fast_sigmoid

Collections.Counter added for python < 2.7

Sigmoid

We were not returning () when the scalar op don't have a version.

it was triing the Image2Neibs.perform method for GpuImage2Neibs.

Do this for amdlibm.

There is code for sigmoid, but it is disabled as it is slower.

this add extra useless code.

Fix-crash in scan grad when compute_test_value=raise

Add test

fix gh-1385

Fix scan gpu loop

[FIX] Py24

update the release instruction to tell to update this page.

Detect if the inner replacement would have a different type,
so that a new gpu_from_host is not inserted in the inner graph
in replacement of the one that was just pushed out.