Generate c{double, float}_fma for pyopencl (#698)

* Generate c{double, float}_fma for pyopencl

* Introduce allow_fp_rerdering option

* Update pyopencl requirement

* Fix is complex

* Update docs on allow_fp_reordering

Co-authored-by: Andreas Klöckner <inform@tiker.net>

* no shortcuts

* add comments as to why we need to generate fma for complex and why not for reals

Co-authored-by: Andreas Klöckner <inform@tiker.net>