spirv-opt: Loop unroller removes RelaxedPrecision decorations

[castano]

When you have a loop that has instructions with RelaxedPrecision decorations, unrolling it maintains the RelaxedPrecision decoration in the first copy of the loop, but removes it from the remaining ones.

For example, in a loop like this:

    [[unroll]] for (uint i = 0; i < 16; i++) {
        mediump float x = inputData[idx * 16 + i] * 16 - sum;
        var += x * x;
    }

The result is as follows:

    float _69 = (_34.inputData[_38] * 16.0) - _523;
    mediump float mp_copy_69 = _69;
    float _102 = (_34.inputData[_338] * 16.0) - _523;
    float _118 = (_34.inputData[_351] * 16.0) - _523;
    float _134 = (_34.inputData[_364] * 16.0) - _523;
    float _150 = (_34.inputData[_377] * 16.0) - _523;
    ...

This results in a significant performance degradation on shaders using mediump/min16float types.

Full example attached: unroll.zip