[Executorch][llama] bug fix for custom sdpa for attention bias

When using attention bias dont override seq length for causal attention

Differential Revision: [D73222733](https://our.internmc.facebook.com/intern/diff/D73222733/)

[ghstack-poisoned]

Author: kimishpatel

extension/llm/custom_ops/op_sdpa.cpp

@@ -400,7 +400,8 @@ Tensor& custom_sdpa_out_impl(
 
   ET_CHECK_MSG(q.dim() == 4, "query must be a 4D tensor");
 
-  const int64_t num_keys_for_causal_attention = start_pos + seq_len;
+  const int64_t num_keys_for_causal_attention =
+      attn_mask.has_value() ? -1 : start_pos + seq_len;
 
   ET_KERNEL_CHECK(
       ctx,