extend design for fpmath mode

rfcs/20240808-graph-api-int-compression-for-sdpa/README.md

@@ -277,14 +277,26 @@ the design of floating point math mode also helps to maintain a consistent API
 semantics between graph API and primitive API. 
 
 Currently, oneDNN Graph API supports setting floating-point math mode during the
-construction of graph object, which will affect the whole graph. The API is
-like:
+construction of graph object, which will affect the whole graph. 
+
+As primitive API has supported a second argument for enforcing an integral
+primitive to comply with the floating-point math mode, oneDNN Graph API need to
+accommodate to the change. The new API will be like:
+
+```cpp
+/// @param mode Floating-point math mode.
+/// @param apply_to_int Use of floating-point arithmetic for integer primitives.
+void set_fpmath_mode(fpmath_mode mode,  bool apply_to_int = false);
+```
+
+Users can use the new API like:
 
 ```cpp
 using namespace dnnl::graph;
 
-// Specify math_mode while constructing the graph
-graph g(kind, math_mode);
+graph g(kind);
+g.set_fpmath_mode( math_mode, /*apply_to_int=*/true);
+
 op foo(id, kind, "foo"); g.add_op(foo);
 op bar(id, kind, "bar"); g.add_op(bar);
 g.finalize();
@@ -297,6 +309,9 @@ cp0.execute(…);
 cp1.execute(…);
 ```
 
+User can still set graph-level fpmath mode through graph constructor, and only
+one method is allowed for a graph.
+
 #### Alternative option
 
 Considering the fact that the framework users can create different graphs if