spirv-opt should be able to optimise two consecutive 'OpVectorTimesScalar's into an 'OpFMul' and a single 'OpVectorTimesScalar'

At the moment, `vec4 * float * float` in glsl compiled via `glslc` becomes:

```spir-v
%12 = OpLoad %v4float %a
%15 = OpLoad %float %b
%16 = OpVectorTimesScalar %v4float %12 %15
%18 = OpLoad %float %c
%19 = OpVectorTimesScalar %v4float %16 %18
```
While `vec4 * (float * float)` becomes:
```
%12 = OpLoad %v4float %a
%15 = OpLoad %float %b
%17 = OpLoad %float %c
%18 = OpFMul %float %15 %17
%19 = OpVectorTimesScalar %v4float %12 %18
```
I'd like `spirv-opt` to be able to do this optimisation for me. Has this been discussed before? Are there any problems with accuracy that could arise from this?

Start with the issue's SPIR-V examples and the spirv-opt entry point, then review how existing optimization passes handle consecutive OpVectorTimesScalar instructions. Determine whether combining the scalar operands preserves the required accuracy, and define completion as a pass that produces the requested OpFMul plus single OpVectorTimesScalar form with coverage for this case.

Written by the indexing model from the issue text.