Add softcap to flash attention

Just CPU and CUDA for now (but, as we know, flash attention
on the CPU is useless in llama.cpp).

On CUDA this improves PP performance quite a bit, especially for
long contexts. E.g., for PP-16384, I now get 3777 t/s.
Without this change, one cannot use FA, and one gets 2300 t/s
(after fusing softcap and softmax), or 2000 t/s without the
fused softcap+softmax.

In comparison, mainline llama.cpp has PP-16384 = 1549 t/s before
PR-8542 (where Johannes Gaessler has also added softcap to FA),
and PP-16384 = 3097 t/s after this PR.

ggml/include/ggml.h

@@ -1811,7 +1811,8 @@ extern "C" {
             struct ggml_tensor  * v,
             struct ggml_tensor  * mask,
             float                 scale,
-            float                 max_bias);
+            float                 max_bias,
+            float                 softcap);
 
     GGML_API void ggml_flash_attn_ext_set_prec(
             struct ggml_tensor * a,