Reactivate transform dialect tests disabled in iree-org#9489. (iree-o…

…rg#9512)

Fixes iree-org#9510.

Merge pull request iree-org#9471 from google/benvanik-tracy-fibers-en…

…able

Adding support for TRACY_FIBERS.

Merge pull request iree-org#9457 from google/benvanik-vulkan-uuid

Allowing deviceUUIDs to be used in Vulkan device paths.

Merge pull request iree-org#9457 from google/benvanik-vulkan-uuid

Allowing deviceUUIDs to be used in Vulkan device paths.

Fixing default CUDA device selection. (iree-org#9454)

Adding an optional workgroup count region to flow.dispatch.workgroups. (

iree-org#9428)

This completes the plumbing as the region is already supported after
outlining.

I hate the syntax, but until we settle on where distribution happens
and potentially split the op into distributed/non-distributed variants
(like the old flow.dispatch.region we had) this will suffice.

Fixing race on semaphore timepoint cancellation. (iree-org#9400)

Fixes iree-org#9392.

Migrate to use upstream bufferization for all the backends. (iree-org…

…#9211)

This commit also removes IREE bufferization pass.

The tests are moved with preprocessing. They are all converted to destination passing style. Some duplicated tests are removed. Some checks are changed because of the upstream issue. Issues are filed.

```
iree-opt \
  --iree-codegen-convert-to-destination-passing-style \
  --canonicalize \
  -cse \
  --linalg-init-tensor-to-alloc-tensor \
  --split-input-file \
linalg_bufferize.mlir
```

Add benchmark models for RISC-V targets (iree-org#9339)

* Add benchmark models for RISC-V targets

The benchmark metrics currently aim for compilation statistics. To run
the full inteference, the cross-compile runtime and the RISC-V
envioronment need to be in CI to run the inferenenc properly.

Also, a more code-size-conscious set of codegen flags should be used
when we have the benchmark infrastructor updated to consume different
benchmark targets more effectively.

Add a pass to check if all ops are vectorized. (iree-org#9260)

The option is disabled by default because not all ops are vectorized in
IREE. This is an useful flag to initially check performance issues.