Renaming the 'iree' dialect to 'util'. (iree-org#6669)

This frees up "iree" for use as the public frontend dialect. The iree dialect was always just a dumping ground for utility ops, so now it has a name to match.

Add new iree-dialects LLVM project. (iree-org#6665)

* Has a single iree_public dialect (with a dummy op).
* Has the boiler-plate for python and CAPI and a test.
* Not yet wired in/building anywhere.
* No bazel build yet.
* Left a number of TODOs for future me to fix upstream.

Convert Einsum to use Matmul/BMM (iree-org#6653)

Error out on unsupported benchmark function input types. (iree-org#6655)

Change tied operand formatting to allow type changes. (iree-org#6631)

The ops have always held the types/dims but they were elided when printing.
Now type and dimension changes can be fully specified and round-tripped.

Reverts iree-org#6188.
Fixes iree-org#6075.
Fixes iree-org#6185.
Fixes iree-org#6420.

Enable fft e2e tests on CUDA backend. (iree-org#6633)

Closes iree-org#6604

[CUDA] Mark argument pointers as align16 (iree-org#6627)

HAL always ensure that buffer are 16 bytes aligned, this is needed to be
able to infer load/store alignment to generate load4/store4

Run canonicalization after tiling during flow dispatch creation (iree…

…-org#6623)

Canonicalization patterns may apply to some ops and cause the root_op
marker to be dropped. Therefore it should be ran after all the tiling
patterns have be ran.

Fix cuda hal for `-Wunused-but-set-parameter` (iree-org#6610)

This is a new check in clang that is going to be enabled globally
internally soon. It's not yet available in a released clang. If it ends
up being annoying, we could disable it for all of IREE internally. We
could also do that now.

Regardless, I think this fix is probably good, since we've identified
the issue. This is currently the only failure in IREE core.

Fix cuda hal for `-Wunused-but-set-parameter` (iree-org#6610)

This is a new check in clang that is going to be enabled globally
internally soon. It's not yet available in a released clang. If it ends
up being annoying, we could disable it for all of IREE internally. We
could also do that now.

Regardless, I think this fix is probably good, since we've identified
the issue. This is currently the only failure in IREE core.