Slow oneAPI performance/stall on simple ML-from-scratch application

A fairly simple ML-from-scratch tutorial, https://github.com/odsl-team/julia-ml-from-scratch, runs fine on GPU and very fast on CUDA.jl, but is extremely slow and stalls at the end on oneAPI.jl.

See odsl-team/julia-ml-from-scratch#2 for performance results.

@pengtu suggested I open an issue here.

PR's to julia-ml-from-scratch are welcome as well of course (maybe there something "oneAPI-unfriendly" in there?).

Start by running the Julia ML-from-scratch tutorial linked in the issue and review the performance results in odsl-team/julia-ml-from-scratch#2. Compare its GPU, CUDA.jl, and oneAPI.jl behavior, including the end-of-run stall. Done means the oneAPI performance or stall has a documented cause and the relevant application or oneAPI.jl behavior is corrected.

Written by the indexing model from the issue text.