Missed optimization: simplify splat-vector bitcast to scalar multiplication/masking

```llvm
define i32 @src(i64 %arg0, i64 %arg1) {
%v0 = icmp eq i64 %arg1, %arg0
%v1 = insertelement <4 x i1> poison, i1 %v0, i64 0
%v2 = shufflevector <4 x i1> %v1, <4 x i1> poison, <4 x i32> zeroinitializer
%v3 = bitcast <4 x i1> %v2 to i4
%v4 = zext i4 %v3 to i32
ret i32 %v4
}

define i32 @tgt(i64 %arg0, i64 %arg1) {
%v0 = icmp eq i64 %arg1, %arg0
%v_i32 = zext i1 %v0 to i32
%result = mul i32 %v_i32, 15
ret i32 %result
}
```

alive2: https://alive2.llvm.org/ce/z/uUn2VN
godbolt: https://godbolt.org/z/3q3fxdYzK
Pattern found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/velox/optimized/Comparisons.ll