FR: autogenerated schedules should try to late-bind natural_vector_size()

[steven-johnson]

Currently, both autoschedulers appear to example Target.natural_vector_size() for the single target passed to them, and generate vectorizations accordingly.

This is fine when targeting a single microarchitecture, but experience has shown that you can get good results across a range of microarchitectures (e.g. x86 SSE4.1/AVX/AVX2) by just vectorizing relative to the specific subtarget in a multitarget build.

Could the generated autoschedule emit vectorizations in terms of what the value of get_target().natural_vector_size<>() is at C++ compiletime? If so, we could probably emit more flexible schedules.

(Yes, this might well constrain other aspects of the schedule, e.g. tiling, so this may be impractical or impossible.)