Optimize M2L for GPU (#138)

* Optimize M2L for GPU

* Move icoeff_tgt to top level iname

* Fix substitution

* use loopy branch

* remove unused imports

* go back to loopy main

* Reduce diff

* move all optimizations to m2l_translation

* Remove extraneous FIXME

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