[regression] ST_EstimatedExtent errors out when no stats are found

[strk]

According to #877 this was fixed 4 years ago, and the PostGIS Manual indeed reports it should return NULL instead since 1.5.4 — ​http://postgis.net/docs/manual-2.2/ST_EstimatedExtent.html

In current trunk, I see it raising an exception instead. It's to be checked WHEN it regressed, and I think we should make it match the documentation…

[strk]

The commit that switched from ERROR to WARNING for 1.5.4: ​https://trac.osgeo.org/postgis/changeset/8319

[strk]

POSTGIS="2.0.8SVN r14008" returns NULL.

Quick testcase:

CREATE TABLE t(g geometry);
SELECT ST_Estimated_Extent('t','g');

[strk]

It's curious that the existing test for ticket #877 is expecting an ERROR instead of a WARNING !

I found the commit that changed such behavior: r10796, by Paul while adding 3d/4d index bindings

[strk]

So it looks like 2.1.0 was the first to switch back to ERROR. What do we do now ? Current situation:

 1.0.0 - 1.5.3     : ERROR
 1.5.4 - 1.5.10    : WARNING
 2.0.0 - 2.0.8dev  : WARNING
 2.1.0 - 2.1.9dev  : ERROR
 2.2.0 - 2.2.1dev  : ERROR

[pramsey]

Going back to warning shouldn't be a problem, it's unlikely anyone depends on error. However, apparently it's sufficiently rare that hardly anyone depends on warning either

[robe]

I've decided shouldn't be done in micro, so will do in 2.3.0

[robe]

In 15096:

Revert ST_EstimatedExtent back to good ol'days when it returned WARNING instead of ERROR when no stats
Closes #3396