* remove usage of parameters in setFromMatrixColumn

* This fixes the usage of setFromMatrixColumn

No vertex and no index for cap with 0 radius

See #8685

The indexCount was rigth but the vertexCount was wrong. The squares on
each plane (face) have common vertices

* MTLLoader: Deprecate setBaseUrl in favor or better named setTexturePath.
Make sure MaterialCreator.baseUrl is not undefined if no path/texture path is set.

* MTLLoader: Automatically resolve texture base path from .mtl source URL if not explicitly set with setPath or setTexturePath.
The URL parameter to .parse(text, url) is passed in if you use .load(). For external use it is optional and preserves earlier behavior if not passed in.

* MTLLoader: Use single quotes consistently. Make params object property setting easier to read.

* MTLLoader: Don't overwrite already found diffuse texture like is done with bump map. This might change behavior for bad materials that define 'map_kd' multiple times.

* MTLLoader: Remaining double quotes to single ticks, also in documentation.

* MTLLoader: Remove auto resolving base path from .mtl source URL.
Add documentation to load and other important functions.
Remove now redundant call to .setBaseUrl in obj/mtl loader example. setPath is enough if the .mtl and textures have the same base path (documented).

* MTLLoader: Dont break parsed texture references that are already absolute URLs.

* MTLLoader: Fix doc typos.

* MTLLoader: Move utility function inside the function that uses it.

* MTLLoader: Use case insensitive regexp instead for abs URL checks.

* Update ISSUE_TEMPLATE.md

* Update ISSUE_TEMPLATE.md

If we use zoom, `camera.fov` is not change. By replacing `camera.fov` by `camera.getEffectiveFOV`, the correct FOV is get.

* ObjLoader: Implement support for multiple materials inside a geometry/object.
- Tracks material declaration occuring mid face declaration
- If object has multiple materials MultiMaterial is creted and geometry groups are created.
- This approach is better than splitting into separate objects: 1) we are creating as many objects the file defines (correctness) 2) The full geometry would be duplicated into
  two objects, afaik this uploads the geometry multiple times to the GPU (perf, i had test asset that had >400 submeshes with 1-6 geometry groups each).

Fixes #8681 #8640
Updates #8203 (fixes OPs 2. point)

* OBJLoader: Code cleanup and three.js style formatting.

* tighten up Vector4.

* tighten up Vector3.

* tighten up Vector2.

* add missing "return this" to Ray3.lookAt().

* tighten up Quaternion.

* missing "return this" on Spherical.

* missing "return this" on Euler.  Tighting up code.

Correct indent.

THREE.SoftwareRenderer alpha flag implementation.

* Update BooleanKeyframeTrack.js

* Update StringKeyframeTrack.js

* merge bufferAttributeTypes with dev

* else if

* serialize the normalized flag

* update docs

* remove support for DataView as BufferAttribute source. This is to specific purpose, and geometry using this will break in too many other methods

* reimplement the DataView as a BufferAttribute data source

* try to warn when bufferAttribute contains a DataView and cannot be read

* Changed too many `warn` to `error`

* Code quality update in BufferAttribute datatype selection

* support load and save of normalized flag in BufferAttribute

* Check for DataView attribute buffers

* mend

* Revert "mend"

This reverts commit 1b4c87687d10fdaefd5007b98019b25d33c492e1.

* try to fix after --ammend causes conflicts

* working on cleaning this all up

* More whitespace cleanup

* more code style cleanup

* strike OBE comment

* more spacing cleanup

* Add a BufferGeometry based Subdivision Modifier

This started from the original subdivision modifier, and works in the
same way except it has different output - and is much MUCH faster and
far easier on RAM. (From >1m and >3gb (queue crashing browsers) to <10
seconds and <500mb on my tests).

Expected input: Index'd Buffer Geometry or THREE.Geometry - with uv's.
Output: Unindex'd Buffer Geometry.

Also introduced is a class I call the 'TypedArrayHelper' which provides
THREE objects (Such as Face, Vec3, Vec2) as 'registers' (I was thinking
like a CPU register), and controls the typed arrays underneath. It
automatically resizes them and provides a trim() function. While
resizing arrays is slow, it's still all way faster than millions of vec3
objects on a regular Geometry.

* Various Fixes

Fixed normal generation according to
https://www.opengl.org/wiki/Calculating_a_Surface_Normal
maybe various other cleanups.

* Revert "Various Fixes"

This reverts commit 4348a5f60b85ad8381e7ed3f19c0602a23bb87c4.

* Revert "Revert "Various Fixes""

This reverts commit c972529a2e0d869ac40d29409beb21420a833ad2.

* change the example to use BufferSubdivisionModifier

It's almost all working except for LatheGeometry. I'm not entirely sure
what is different about the lathegeometry , but I will investigate later
unless someone else figures it out first.

* change back to v2 and use phongmaterial

change back to v2 and use phongmaterial

* dispose of geometries

see title.

* Remove typedarray.slice's

in favor of typedarray.subarray's.

* slight cleanup to subarray

simply set the new buffers contents to the old buffers contents if
possible, otherwise use subarrays

* Tightened flattening of uniform array data.

* Updated builds.