Is xfb_stride must be multiple of 4

I use int8 or int16 to test a transform feedback shader like this:

layout(location = 0, xfb_buffer=0, xfb_stride = 2, xfb_offset = 30) out uint16_t u16;

An error is reported concerning xfb_stride incorrectness.

Apparently, the xfb stride computation does not take 8-bit/16-bit into account because the standard xfb does not support such data types. But we do have 8-bit/16-bit types introduced by EXT_shader_explicit_arithmetic_types. For AMD, 16-bit data exported to xfb buffer is allowed by GL_AMD_gpu_shader_half_float. So I wonder if we can introduce such computation as a general solution. If you have concerns, I could separate the codes with AMD_EXTENSIONS.

Start by locating glslang's transform-feedback validation for xfb_stride and compare its rules with EXT_shader_explicit_arithmetic_types and GL_AMD_gpu_shader_half_float. Use the reported uint16_t shader with xfb_stride=2 and xfb_offset=30 as the reproduction; done means the validator applies the appropriate 8-bit/16-bit rules or clearly rejects unsupported cases.