[TritonNVIDIAGPU] Add MemRead<GlobalMemory> to async TMA write ops (triton-lang#6485)

Author: Mogball

include/triton/Dialect/TritonNvidiaGPU/IR/TritonNvidiaGPUOps.td

@@ -315,7 +315,7 @@ def TTNG_AsyncTMACopyLocalToGlobalOp : TTNG_Op<"async_tma_copy_local_to_global">
   }];
 
   let arguments = (ins
-    Arg<TT_PtrType, "", [MemWrite<GlobalMemory>]>:$desc_ptr,
+    Arg<TT_PtrType, "", [MemRead<GlobalMemory>, MemWrite<GlobalMemory>]>:$desc_ptr,
     Variadic<I32>:$coord,
     Arg<TTG_MemDescType, "", [MemRead<SharedMemory>]>:$src
   );
@@ -364,7 +364,7 @@ def TTNG_AsyncTMAScatterOp : TTNG_Op<"async_tma_scatter"> {
   }];
 
   let arguments = (ins
-    Arg<TT_PtrType ,"", [MemWrite<GlobalMemory>]>:$desc_ptr,
+    Arg<TT_PtrType ,"", [MemRead<GlobalMemory>, MemWrite<GlobalMemory>]>:$desc_ptr,
     RankedTensorOf<[I32]>:$x_offsets,
     I32:$y_offset,
     Arg<TTG_MemDescType, "", [MemRead<SharedMemory>]>:$src