Actually, I think this can be handled magically instead of requiring understanding of the concept that the camera is part of the scenegraph.
While I was at it, I fixed it so you can use a camera for rendering different scenes (the renderer will add it and remove from each scene graph)	.

Removed LOD text example. I think LOD example is good enough.
Ubiquity test working again.
Tweaked README to show the new pattern (Scene first thing created).

This doesn't handle properly order of updates of LOD vs camera.

Code is a horrible mess but at least it works. Anyways, some day we should revisit GPU accelerated skinning with a proper pipeline.

I had to resurrect parts of old code (keeping track of scene graph changes).

Good news: performance is great, new scene graph gains compound with buffers optimizations and non-array materials refactoring. Together we gained about 15% for rotating cubes performance test (pushing max to 2650 cubes at 60 fps).

Still broken with new scene graph: lod, skinning, collisions, AnaglyphWebGLRenderer, ParallaxBarrierWebGLRenderer, multi-canvas circle example.

Also moved `.overdraw` to Material.
Code is much simpler now :)

This is to be able to skip redundant updates in multipass rendering (like for shadow maps or dynamic cube maps).

Also reshuffled initWebGLObjects calls to have just a single update in a basic scenario of shadow map pass followed by scene render.

Still need to solve somehow potential redundant updates in EffectComposer (which can mix multiple scenes).

Just for the sake of consistency, when object properties are not involved performance difference is negligible.

These little buggers were costing 3.35% of the total time in performance test.

Noticeable difference: can't anymore assign Material references to faces directly - "face.materials" became "face.materialIndex". Materials have to go to "geometry.materials" array (to which "face.materialIndex" points). This is necessary to allow run-time changing of face materials (this indirection used to be done by arrays).

Some casualties:

- multi-materials example, this functionality isn't possible anymore
- had to simplify LOD Text examples as LODs don't handle groups (used to fake multi-materials)
- multi-materials sphere in materials example can't have faces with undefined materials (this may be fixable I guess, when there'll be some real use-case)
- scene format still has materials arrays, only the last material will be used (except when there are face materials, then materials from JSON will be used, as before)

Other renderers and their corresponding examples were not touched yet.

WebGLRenderer still needs some prettification, some things don't make sense anymore without arrays, code can be cleaned up further.

New per-pixel Phong code is enabled by setting `perPixel` parameter in Phong material.

This version is able to light single untesselated quad correctly.

Also it doesn't have point lights number limitation caused by varyings, all lights pass just through uniforms.

Performance wise it seems surprisingly quite similar to old version (stress test is ~6 fps faster with new version, though I guess for different use cases / GPUs it can differ).

Metal makes specular term multiplicative (default is additive). This is to be able to have textured materials that do not have diffuse color, just a specular one, to get metallic look (additive specular gives plastic look).

Removed diffuse color influence on ambient term in Phong. This was causing texture disappearance when diffuse color was black.

Made Lambert more similar to Phong: now it also has ambient color parameter. Before ambient light was affecting Lambert as if ambient color was always 0xffffff.

This is not supposed to be "proper" solution (for example this doesn't solve removing of objects), but at least some WebGL examples are working now.

Lots of code got cleaner with seconds, seems we used them before all around.

Also changed examples that didn't need deltas to simple Date.now() plus some more examples cleaning, trying to get rid of obsolescence warnings.

Apparently iterating over object properties is dangerous even if that code path is not executed.

Stress-testing particles buffers setting showed 60% performance increase (in Chrome) even when there were no custom attributes (and thus custom attributes handling code behind "if ( customAttributes )" was never executed).

I suspect this may be caused by some V8 optimization heuristics.

This is for being able to render shadowmap just once per frame - all dynamic cubemap faces and full scene can share the same shadowmap (at least until we don't have camera-specific fine-tuned shadowmaps).

Added morph to rendering testbed.
Simplified shadowmap shader snippet gamma handling.

Do not merge, this is very experimental, not sure about correctness, also tons of unsolved things.

Turns out before all cube texture faces shared a single depth buffer.

Last cube face still missing.

Before lighting was messed up in dynamically generated cube maps.

Dynamic cube maps need to have flip uniforms (flipEnvMap / tFlip) set to 1, static cube maps to -1 (default).

For standard materials this is handled automatically, it's just something to be aware of when using cube maps in custom ShaderMaterials (dynamic and static cube maps have different cube face orientations, dynamic ones are correct, static ones need flipped x-axis in lookup vector).

There is still mystery of disappearing last cube face (negative-Z) and weirdly clipped objects in cars demo (works ok in simpler test).

Now `Camera` can use inherited `update()`

This feels more intuitive, otherwise ambiently lit textured surfaces look weird.

Turns out I had it right before after all: to get rid of artefacts it should be multiplied by diffuse term (instead of being patched by conditional).