Hexagon conv2d full output slice (apache#9198)

* split h axis by constant factor 2; no cache write

* enable cache_write, but not yet able to compute_at

* cache_write with compute_at

* cleanup, make loop split semantics more clear

* parameterize height loop split

* nhwhwc wiggling (needs cleanup)

* added input channel splits for crouton depth

* cleanup variable names and magic numbers

* comments

* add README

* added 3x3 conv2d (no padding) case

* add ASF header and RFC link

* cleanup README