ml: only use flash attention on gpu by default

* env VISP_FLASH_ATTENTION=0 always disables it
* env VISP_FLASH_ATTENTION=1 always enabled it
* all other values use default