GitLab

- Add --host-pipeline=runner to rocmlir-driver, which does the
[gpu and a bunch of stdlib dialects]->llvm steps to make host code and
which compiles gpu kernels to HSA code objects
- Make -c mean "-kernel-pipeline full -host-pipeline runner" instead
of "-kernel-pipeline gpu"
- This means that (with the exception of xmir-runner, which needed to
do some runtime stuff to pick the target), we no longer have "runners"
that do a bunch of non-trivial work on their input. An advantage of
this is that benchmarking now doesn't include the time it takes to
compile a GPU kernel (which it might have before)
- Adjust the xmir runner pipeline to include the arithmetic ops
expansion pass that was needed for host code validation, and rearrange
it some to unify it with the old mlir-rocm-runner
- Delete mlir-rocm-runner
- Adjust tests, mainly
  - Switching all instances of mlir-rocm-runner to mlir-cpu-runer -O2
  - Change the single-target verifier=clone tests to use the runner
    pipeline because they don't need the xmir functionality
  - Change the convolution harness tests to expect -c to produce a
    binary blob and not a gpu.module
  - Change various CPU e2e tests to include the relevant
    --convert-*-to-llvm invocations or to go through the runner
    pipeline
- Change the CI perf/parameter sweep scripts accordingly
- Change the rocm-run widget to pick up mlir-cpu-runner