!8465 [MSLITE] mnet full quant bug， perchannel inputsum bug

From: @ling_qiao_min
Reviewed-by: @HilbertDavid,@hangangqiang
Signed-off-by: @HilbertDavid

mindspore/lite/nnacl/assembly/arm64/PreSum4x16Int8Peroc.S

@@ -104,11 +104,11 @@ OcRes4_1:
   b OcRes4End
 
 OcRes4_2:
-  ld1 {v15.h}[0], [x16]
+  ld1 {v15.d}[0], [x16]
   b OcRes4End
 
 OcRes4_3:
-  ld1 {v15.h}[0], [x16]
+  ld1 {v15.d}[0], [x16]
   add x16, x16, #8
   ld1 {v15.s}[2], [x16]
   b OcRes4End

mindspore/lite/nnacl/int8/conv_int8.c

@@ -406,12 +406,12 @@ void Conv1x1PreOptPeroc(const int8_t *src_input, int8_t *packed_input, int32_t *
 
       "14: \n"
       "ld1 {v16.4s}, [x10], #16\n"
-      "ld1 {v17.h}[0], [x10] \n"
+      "ld1 {v17.d}[0], [x10] \n"
       "b 16f \n"
 
       "15: \n"
       "ld1 {v16.4s}, [x10], #16\n"
-      "ld1 {v17.h}[0], [x10] \n"
+      "ld1 {v17.d}[0], [x10] \n"
       "add x10, x10, #8 \n"
       "ld1 {v17.s}[2], [x10] \n"
       "b 16f \n"