[aarch64] Missed MADD folding opportunity

For the following function foo,

```c
long foo(long stride, int mvx, int mvy)
{
return (long)(mvy >> 2) * stride + (mvx >> 2);
}
```

LLVM currently generates MUL+ADD, instead of MADD.

```asm
foo:
sbfx x8, x2, #2, #30
asr w9, w1, #2
mul x8, x0, x8
add x0, x8, w9, sxtw
ret
```

With MADD, we could generate the following:

```asm
foo:
sbfx x8, x2, #2, #30
sbfx x9, x1, #2, #30
madd x0, x0, x8, x9
```

The test case was found in DCPerf's VideoTranscodeBench benchmark.