Disabling ggml_flash_attn_ext_set_prec

[pt13762104]

llama.cpp/src/llama-graph.cpp

Line 1308 in 3c3635d

ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);

Does disabling this have any implications? Without this, the FP16 kernel gets used, it gets about 5-10% more performance in PP.

llama.cpp/ggml/src/ggml-cuda/fattn.cu

Lines 414 to 415 in 3c3635d

	if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
	return BEST_FATTN_KERNEL_TILE_F16;

(I tested GPQA with this disabled and the results looks fine)

[pt13762104]

since #15769, this is a no-op.