llama.cpp prefill performance drops on ROCm

While token generation improvements are a plus, the regression in prefill performance is a critical blow for users relying on fast initial response times in multi-GPU setups. ROCm's optimization path appears to favor throughput over latency in recent commits, potentially alienating users with complex workflows. Furthermore, the shift in optimal unified block sizes suggests underlying changes in memory management on RDNA 3 hardware, raising concerns about PCIe bandwidth efficiency for dual-GPU configurations.