Instead of waiting for all materials of a mesh before fetching geometry
data, we now push all promises into one large array and wait on all of
this at once.

This means we can load .bin file in parallel with texture data when
texture assets are using separate files, and potentially start decoding
geometry earlier.

When used from GLTF loader, it's common for BasisTextureLoader promises
to be created in quick succession for a number of textures.

What happens is that the methods wait until the transcoder is loaded,
and then call _getWorker in succession. After each _getWorker call
succeeds, the continuation for the promise is scheduled - however, the
worker task cost is only adjusted when that continuation is ran, and the
next _getWorker call runs immediately after.

As a result, the _taskLoad-based sorting doesn't work, and most textures
are assigned to the last worker, which means that the decode is mostly
single-threaded.

This change fixes this by integrating cost adjustment into _getWorker
(renamed to _allocateWorker for clarity); we also don't need _taskCosts
anymore because we can close over the taskCost local instead.

Also fix the problem where we were creating redundant object URLs for
the transcoder binary because of many calls to _initTranscoder (during
_getWorker) during which the transcoder promise is setup but the binary
hasn't been loaded yet.

Almost all models that use occlusionTexture use texCoord 0 - the loader
prints a console message for each material for such models which seems
redundant / excessive.

However, if other textures use texCoord that's not 0, the resulting
texture will not be rendered properly. We now detect this case and issue
a warning.

This fix creates all uv channels, not only the first one

- Removes GeometryUtils random point selection.
- Adds MeshSurfaceSampler for incremental, weighted surface sampling.
- Adds 'webgl / instancing / scatter' example with InstancedMesh.
- Adds 'Flower.glb' sample for the example.