[PIPELINE] AssignLatencies for mmav5 (#6077)

Implement and test assigning latencies for mmav5 ops (regular and
scaled). This PR also features slight refactor of AssignLatencies code
with putting load-related and mma-related code in separate classes for
maintainability.
Latency assignment for mma ops is enabled only for tests now, disabled
in normal flow, as scheduling and lowering is not implemented yet.

---------

Co-authored-by: Thomas Raoux <[email protected]>

Author: pawelszczerbuk

include/triton/Dialect/TritonGPU/Transforms/Schedule.h

@@ -15,7 +15,7 @@ namespace gpu {
 
 /// Discover operations that should become async and assign latencies to them
 /// based on the numStages value provided by the user.
-void assignLatencies(ModuleOp moduleOp, int numStages);
+void assignLatencies(ModuleOp moduleOp, int numStages, bool assignMMA = false);
 
 /// Schedule the loops based on the latencies assigned to the operations.
 void scheduleLoops(ModuleOp moduleOp);