Instead of changing the extension parsing loop, we now use the fallback
only if extension is not present in the required list.

If KTX2 loader wasn't registered, we would previously fail to load GLTF
files that used BasisU; this is a problem since files that don't require
BasisU support but specify it optionally would incorrectly fail.

The tradeoff here is that the error message that we previously emitted
for cases where the extension was required is now no more; we need some
mechanism for required extensions to actually be enforced to solve this.

Long term we might want to have a way to indicate failure from a plugin,
but for now this seems reasonable.

This makes sure that files with EXT_meshopt_compression used as a
non-required extension don't fail to load.

EXT_meshopt_compression is a recent extension that supports compressing
glTF data - geometry, animation, instance transforms - in a generic
fashion.

Specification:
https://github.com/KhronosGroup/glTF/tree/master/extensions/2.0/Vendor/EXT_meshopt_compression

It is possible to produce files that use the extension using gltfpack
(https://github.com/zeux/meshoptimizer/tree/master/gltf). To decompress
the files, loader needs to decode the bufferView data - the extension
specification details the compressed format, but meshoptimizer library
provides a portable decoder that can be readily consumed from JS; the
decoder requires WebAssembly and will automatically dispatch to Wasm SIMD
when available, or use MVP otherwise.

meshopt_decoder.js is a drop-in copy of
https://github.com/zeux/meshoptimizer/blob/master/js/meshopt_decoder.js;
since it implements the specified format, it can be updated at any point
without loss of backwards compatibility.

The PR for EXT_meshopt_compression in glTF repository
(https://github.com/KhronosGroup/glTF/pull/1830) contains file sizes for
a lot of example models - the benefit of using this extension is that it
supports compressing every possible type of glTF binary data except for
images, and can decode at breakneck speeds (~1 GB/s when using Wasm SIMD
on a recent desktop Intel CPU) while only requiring a very small decoder
binary (meshopt_decoder.js is at 21 KB raw, 6 KB deflated, while
including *two* builds of the decoder, scalar and SIMD).

As an example, I've tested this PR on BrainStem model from
glTF-Sample-Models repository that contains animation data. The original
file size is 3120 KB, Draco-compressed version is 1942 KB, and
meshopt-compressed version is 362 KB, produced by gltfpack with `-cc`
option. The large size delta is due to animation data being compressed
with the extension; the loader extension decompresses bufferViews upon
request and the rest of the loader pipeline works without further
changes.

To integrate the extension after this change the user simply needs to
include the module, and pass the exported MeshoptDecoder global to the
loader:

	import './js/libs/meshopt_decoder.js';

	...

	loader.setMeshoptDecoder( MeshoptDecoder );