[glsl-in] Incorrect array index calculation

[magcius]

Description
naga uses an i32() cast to convert from the floating-point array index specified in the texcoord vector to an integer array texture. However, the GL specification says that round-nearest-even should be used: https://registry.khronos.org/OpenGL/specs/gl/glspec46.core.pdf#page=281

Repro steps
This is visible in the output of samplers.frag.wgsl in the test suite:

wgpu/naga/tests/out/wgsl/samplers.frag.wgsl

Lines 415 to 416 in c54a159

	let _e34 = coord_9;
	let _e38 = textureSample(tex2DArray, samp, _e34.xy, i32(_e34.z));