Co-authored-by: Valentin Churavy <v.churavy@gmail.com>
(cherry picked from commit 474050e)

* Use Float32 in examples with backends that don't support Float64

* Add `backend` argument for `examples_testset`

* Reduce `TILE_DIM` for compatibility

Metal doesn't always support 1-24 threads, which causes intermittent errors with 32x32 tiles

* Fix histogram implementation

The final part of the loop expects every thread to exists, so we cannot not launch them. Avoid work on extra threads until then.

Also use Int32 since some backends lack Int64 atomics, and make one of the tests have weird groupsize since that's when the errors used to pop up.

(cherry picked from commit dab03b9)