vulkan: For coopmat2 FA, use fp16 accumulators for the final result (#19376)

The cpu and cuda backends use fp16 for the VKQ accumulator type, this change
does the same for vulkan. This helps particularly with large head sizes which
are very register-limited.

I tried this for the coopmat1 path and it slowed down a bit. I didn't try for
scalar.

I applied the softmax bias that the cuda backend uses to avoid overflow,
although I was not able to reproduce the original bug without it.

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl

@@ -240,3 +240,7 @@ void init_indices()
     // and breaking the alignment detection.
     m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
 }
+
+// Bias applied to softmax to stay in fp16 range.
+// Based on ggml-cuda issue https://github.com/ggml-org/llama.cpp/issues/18606
+const float FATTN_KQ_MAX_OFFSET = 3.0f*0.6931f;