fix crash when using dtype=float16 (assigned tuple instead of integer).

Fix indexing with reverse slices

An earlier commit fixed a rounding issue for slices with steps > 1, but it broke reverse indexing, where step < 0.  This fix should re-enable reverse indexing.

Fixes to enqueue_copy_buffer_rect

There are two bugs:
1. If src_pitches or dst_pitches is specified, there is an uninitialised
   variable.
2. If ConstBuffer is given a short array and content==0, the internal
   buffer is not zero-filled, causing garbage to be passed on.

The latter probably affects a lot of other functions, but hopefully in a
positive way.

Fixes #105.

Revert "Add #ifdef gate around CL_KERNEL_ARG_TYPE_CONST (reported on missing on some Windows ICD by Christoph Gohlke)"

This reverts commit 7a9da07c.

Add #ifdef gate around CL_KERNEL_ARG_TYPE_CONST (reported on missing on some Windows ICD by Christoph Gohlke)

This reverts commit 9739700c.

This reverts commit ca5660c4.