Feature Request: IBM Granite 4.1 models.

IBM's Granite 4.1 models are by far the best local AI models I've ever used (or, at least, the best models within the 1-9 billion parameter range). They lack reasoning capability yet somehow far outperform models such as those in the gemma4 series in reasoning tasks. A good example is the following:

"I have a strawberry, a cup, a table and a microwave. I put the strawberry in the cup, then put the cup on the table. I then turn the cup upside down on the table, and then put the cup in the microwave. Where is the strawberry now?"

I consistently got an answer along the lines of "The strawberry is in the microwave." from competitors (Phi4, Gemma4), often accompanied by minutes-long reasoning strings (on CPU). Granite gave me the correct answer (The strawberry is on the table) every time, with no reasoning involved. It is also good to note that Granite is very good at mathematical tasks, and is ranked as the world's most transparent AI model. For these reasons and others I believe that it is worth getting these AI models working on FLM, that I may use them daily with my potato NPU.

The issue names no files, tests, or entry points. Start by locating existing model integrations; the repository mentions DeepSeek and Llama, so compare their support paths before assessing IBM Granite 4.1 on AMD Ryzen AI NPUs. Done means Granite 4.1 models load and run successfully in FastFlowLM.