While the last commit prevented a race condition, it also broke the
`updateOnce` function.

Thanks to @strarsis.

This is a result of browsers' stricter protections around accessing device
sensors.

This was an issue in Chrome >= 70.

Thanks to @KUJensen for suggesting the subsequent-sibling combinator as a fix.

As iPads use the same user agent as Mac, we need to check if touch input is
supported instead.

This allows interoperability with other WebGL code.

If the viewer is destroyed, we now try to abort loading equirectangular and
cube map panoramas. This appears to work in Chrome but not in Firefox.

No attempt is made to stop loading multires images, since that process is more
complicated.

Previously, a single multires tile level was used for the entire viewport,
which resulted in too much resolution at the center and too little at the
edges. This was particularly noticable when zoomed out all the way.

The new approach selectively loads tiles based on their individual projected
resolution, which is more efficient and should resolve long-standing
edge-softness issues.

It was added in 950942830fb531bf8faee19f85dea5cc94a23de1, but needs more work.

Also add ability to provide a hook to execute just prior to rendering.

It doesn't make much sense to provide a URL to a non-WebGL viewer anymore.

Regression caused rendering issue for cubemaps (but not multires) when CSS
fallback render was used. The regression was introduced as an optimization in
5eab5366f6.

Bug was introduced in 5eab5366f6. It caused tile
loader objects of non-web-worker-based tile loading method to sometimes be
lost during scene changes.

This probably also speeds up the calculation.

This regression was introduced in 9788fc6512.

WebGL 1 doesn't support REPEAT clamp mode for non-power-of-two images, which
is needed to avoid some rendering artifacts for small previews.

Spherical harmonic transform (SHT) previews are display instead of a solid
background color before the base multires tiles are loaded. If a thumbnail is
provided, it will be displayed instead of the SHT preview. If the thumbnail is
provided as a Base64-encoded string, the SHT preview may be briefly displayed,
since image loading is asynchronous, while the SHT preview is loaded
synchronously.

I'm using an 3rd party tool to produce the tiles for the faces - it offers 4 patterns for file and folder hirarchy/naming but all start at level 0.
This is an easy fix that may come handy for others as well.

This reverts commit 7362f5c948.

The bug in question was actually already fixed in
5f0a1d8662.

* 64c337b

* Update build.

* Build 05837fd.

* Build d1555b3,

* Build eb41687.

* Build b232a6b.

* Build 233b100.

* Build ab7b859.

* Build 805f6da.

* Build 25e3d6d.

* Build 54a8dc6.

* Build 0d32eea.

* Build eeb9383.

* Add code.

* Improve/fix interaction message fade.
Fix indentations.

* Build for UI confirmation (40ffc37).

* Ignore build/ again.

* Change interation message text and style.

* Don't display interaction messages while in fullscreen.

* Don't change defaults and add documentation for interation messages.

* Disable `ctrl` / `shift` zoom when `ctrl` is required for mouse wheel zoom.

* Show interaction key based on platform.

* Cleanup.

* Remove pre-built files.

* Use `<div>` for interaction message.

* Fix case without drag confirm and show proper message for `both`.

Also switch back to using separate `dragConfirm` parameter.

* Document interaction message events.

Co-authored-by: Matthew Petroff <matthew@mpetroff.net>

This reverts commit ab423f66ea.