Time for upstreaming

[IMbackK]

Hi hjc4869,

I more or less now maintain the hip backed of llamacpp now. After some discussion with Johannes Gaessler we have now come to the point that it is likely that i will be allowed to merge this into upstream llamacpp. Since you have been rebaseing this regularly i find it correct to let you do the honors.

I must request some changes:

1. Update your patch to add an option GGML_HIP_ROCWMMA_FATTN here which is off by default
   • ifdef away the rocwmma flash attn path when this is not set, including the include of rocwmma
2. When GGML_HIP_ROCWMMA_FATTN is on, use cmake's CHECK_INCLUDE_FILE to check for the rocwmma header and abort if its not found.

Please then open a pr with this against llamacpp and request myself and Johannes reviewers.