Disentangle nightly and simd_support feature

[vks]

Now nightly can be used without implying simd_support. This
hopefully makes the nightly feature more stable and improves
compilation time if simd_support is not required.

The documentation and changelog were updated accordingly.

Fixes #1047.

[vks]

This will also require an update to the Rand book.

The CI does not check whether nightly and simd_support work independently. Should we add this? This would require running the tests two more times, or alternatively running cargo check two times.

[vks]

It seems like __m64 was removed from nightly, so maybe we should stop supporting it as well. But maybe that can wait until packed_simd is fixed.

[dhardy]

LGTM. Yes, ideally we should have at least one test of each independently. We could just use different flags on Travis and AppVeyor.