xrt: Add at v2.21.75

[GZGavinZhao]

Summary

Fixes #9262.

Now people can run ML models (mostly LLMs these days) on their Ryzen AI processors' NPU without eating up GPU resources (i.e. those AMD processors with NPUs, all the Ryzen AI XXX models).

Test Plan

• xrt-smi examine shows my NPU correctly.
• Built FastFlowLM from source and ran Qwen3-8B, producing reasonable results at reasonable performance.

Checklist

• Package was built and tested against unstable
• This change could gainfully be listed in the weekly sync notes once merged
• I agree to license this contribution and all my previous contributions under the licensing terms in LICENSE.md and have the power and authority to grant those licenses.

[GZGavinZhao]

Investigating static library shenanigans...

[EbonJaeger]

I wonder if either the library portion or the /usr/bin portion should be split into a separate package. How is this meant to be used?

[GZGavinZhao]

Thanks for the feedback, will address them later today.

[GZGavinZhao]

Usually downstream consumers (e.g. LLM runtimes like FastFlowLM) link against the dynamic library, However, if you're using the NPU, you almost always want the command-line tools like xrt-smi as well to inspect and manage the state of the NPU. Therefore, I'm inclined to not split the libraries and binaries.