[ci] add rocm build to pull (#80149)
these don't use the AMD runners so we can have as many as we want
without affecting queue times. Would be good to get early signal on rocm
build failures.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80149
Approved by: https://github.com/jeffdaily, https://github.com/seemethere, https://github.com/malfet
