ggml: Don't assert fail when tensor data changes (ggml-org#13222)

The following scenario will cause an assertion failure in the graph
allocator:
 - Build and allocate a graph containing a tensor with a non-NULL data
   pointer
 - Build and allocate a new graph where that data is NULL

Result:
ggml-alloc.c:819: GGML_ASSERT(talloc->buffer_id >= 0) failed

This happens during revalidation because we think that memory should
have been previously allocated based on the current graph but in
reality the previous graph was different. In this situation, we
should do a full reallocation pass.

Author: jessegross

ggml/src/ggml-alloc.c

@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
     size_t node_size = 0;
     if (!node->data && !node->view_src) {
-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
         node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     }
     return talloc->size_max >= node_size;