GLES2 fix octahedral half float unpacking

[lawnjelly]

The strides in _unpack_half_floats() were incorrectly calculated in the case where octahedral normals and tangents were in use.

Notes

• I discovered this while debugging a crash in my BGFX rasterizer, which copies the unpacking code from GLES2.
• The bug occurs because the strides are incremented at the end of the loop:

		src_stride += src_size[i];
		dst_stride += dst_size[i];

Whereas the code in question modifies the size of src_size[VS::ARRAY_NORMAL] rather than the current attribute, therefore the stride is never added.

• This could do with being independently checked by e.g. @clayjohn
• I'm not sure altogether whether this bug occurs in the wild (maybe this combination does not occur usually in GLES2) but it seems worth closing.

[akien-mga]

Thanks!

[norrath-hero-cn]

BGFX rasterizer? this just like a dream comes to true! you make Godot 3.x even greater!!!

[akien-mga]

Cherry-picked for 3.5.2.