Fix Iris matrices when rendering sublevels

[djefrey]

Iris replaces ShaderInstance with ExtendedShader.
When ExtendedShader.apply() is called, modelViewInverse and normalMatrix are updated using MODEL_VIEW_MATRIX value.

However, in VanillaChunkedSubLevelRenderData.renderChunkedSubLevel(), MODEL_VIEW_MATRIX.set() is called after apply() has already been called. This leaves the Iris matrices in an incorrect state.

This can be fixed by delaying the apply() call until just before rendering.

Before:

After:

[ryanhcode]

tested, awesome! What's with the dynamic effects uniforms not needing to be uploaded anymore?

[ryanhcode]

oh, is it because the apply is afterwards now and uploads them anyway? yeah okay

[djefrey]

I made the change when I removed the now useless uploads. However, as I'm reading the code again, I'm not sure if this is actually correct.

[ryanhcode]

will this have any more noticeable impact than just uploading the individual uniforms like we were doing before?

[djefrey]

I've done a simple benchmark: 354 sublevels on a flat world, no shaderpack.
This commit: ~760fps
With uploads readded: ~725fps
So it seems there's a small gain to prevent unnecessary uploads.

However, I noticed that this commit introduces a significant performance regression, as the previous version is around ~1400fps. This regression is caused by apply() reupload every uniforms, as the dirty flag in Uniform is ignored.
The only fix I see would be to compute and update manually modelViewInverse and normalMatrix if shader is an ExtendedShader instance.

[ryanhcode]

With those numbers, I think it's worth it to have explicit compatibility with Iris here