[Architecture request] moonshotai/Kimi-Linear-48B-A3B-Instruct

[robfuscator]

Any chance we could get support for moonshotai/Kimi-Linear-48B-A3B-Instruct? It looks very promising and could neatly fit in 48GB VRAM with enough context at a decent BPW.

I'm not capable of implementing support myself, but will gladly assist in testing, if that's helpful.

Thank you for your work, it is highly appreciated.