Support a relaxed casting mode

Currently, if you want to set any limited-precision number in any situation, you have to cast.

```
uint16_t val = uint16_t(0);
const uint16_t arr[4] = { uint16_t(0), uint16_t(0), uint16_t(0), uint16_t(0) };
```

This leads to code interacting with mixed-precision numbers requiring casts in every single line, such as throughout this [entire file](https://github.com/FFmpeg/FFmpeg/blob/c3d3377fe14645dbb63d937c45df12f3f6746d17/libavcodec/vulkan/rangecoder.comp#L49).

Everything here was designed to not overflow. Generally programmers know whether something will overflow or not.
Having a `#pragma releaxed_casting` would be a good quality of life improvement.

Start by reading the referenced FFmpeg file, libavcodec/vulkan/rangecoder.comp, to understand the repeated limited-precision casts motivating this request. Then trace glslang's existing pragma handling and tests; done requires a defined relaxed-casting mode with documented behavior and coverage for the mixed-precision cases described here.

Written by the indexing model from the issue text.