[LLVMGPU][ROCm] Add I8 MFMA layout support for CDNA2

[nirvedhmeshram]

These i8 intrinsics are supported by CDNA2 arch so adding them provides better functionality such as fixing #18406. It is worth noting that, this i8 support is emulated and the CDNA2 hardware does fp16 under the hood for these intrinsics.
With this PR we can also support: "sharktank/punet/int8" from the external test suite and hence the xfail on those tests are removed.
Fixes: #18406

[nirvedhmeshram]

I considered adding matmul e2e tests for this, but couldnt see a way short of writing a new generate script that will allow us to run the suite for cdna2 without having instrinic mismatch with cdna3. Given that we are testing some models e2e for cdna2, I am okay with this but just wanted to call it out in case reviewers have other idea.

[ScottTodd]

I considered adding matmul e2e tests for this, but couldnt see a way short of writing a new generate script that will allow us to run the suite for cdna2 without having instrinic mismatch with cdna3. Given that we are testing some models e2e for cdna2, I am okay with this but just wanted to call it out in case reviewers have other idea.

I'm in the process of migrating the matmul e2e tests to https://github.com/iree-org/iree-test-suites/tree/main/linalg_ops and refactoring to be decoupled from the core IREE CMake project (see iree-org/iree-test-suites#2). Next on my list is refactoring the massive list of iree_generated_e2e_runner_test test suites. Can add cdna2 support to the list.

[ScottTodd]

Note that in the new location, the tests are not using iree_codegen.compilation_info at all. The compiler must determine which pipeline to use based on the target device and other user-facing flags, not explicitly added IR.