Just wasted half an hour trying to track down a crash in sc unit test
that was apparently caused by different object files being compiled
with different OSL_DEBUG_LEVELs and thus some having XclDebugObjCounter
and some not.

- instead on lots of recursive build.pl modules to iterate through we build all
  tail_build up to the requested module inside one tail_build when doing a
  "make foo.all"
- advantage: speeding up because getting rid of lots of recursive makes with
  artificial module barriers, forcing almost sequential execution: build, link,
  unittest with lots of cpu-idling at each module-end
- disadvantage: all shared dependencies of tail_build are now build for a
  (sub)module of tail_build. This is transitional as we migrate more stuff into
  tail_build and there are good dependencies inside tail_build
- "make foo.all" now only executes slowcheck now for foo, not for all the deps

Or maybe it's the Mach-O object file format that lacks the
functionality? Anyway, let's try to manage without visibility then.

For unrecognized options, Clang 3.1 (and ealier?) prints out a
warning, but it isn't a real warning as it doesn't cause a compilation
error with -Werror, so it is impossible to find out reliably if an
option is supported or not. So hardcode the information we know.

Although Microsoft says that SelectionTree control can publish a control
event only on Windows Server 2003 and above, the  custom action seems to
be working under a fully patched Windows XP SP3. Maybe it fails silently
on older Windows XPs, not to mention Windows 2000. I did not test those.

...as checking the lastsynchronized time-stamps can give wrong results if new and old
impl use same user config and old impl is started once more after new impl is installed,
in which case per-user data of bundled exts of new impl would not be regenerated,
leading to inconsistencies.

...otherwise, if they are bundled, their per-user data is not regenerated, leading
to inconsitencies.

dbaccess_complex segfaults because ptr_vector::erase actually deletes
the element; also, a std::vector is more appropriate to store smart ptr.
(regression from 8533a191)

has been return SYMBOL since initial checkin in 2000,
SYMBOL is a non-0 enum so return true here instead

This reverts commit cf9c715a.

3c82780cbc7cf17cf348e7f49b99663cfdc42ed7 "Set gb_DEBUGLEVEL=1 for
--enable-dbgutil after all" fixes the underlying problem, making this awkward
workaround no longer necessary.

This is a partial revert of db8df57a "fixes for
debug flags handling."  Removing this caused various nastinesses, at least:

* Some tinderbox started to emit bogus compiler warnings at -O2 vs. -O0, see
  cf9c715a "Avoid bogus warnings (GCC 4.4.6)"

* General DBG_UTIL vs. OSL_DEBUG_LEVEL confusion in the code, see
  729165eb "OSL_DEBUG_LEVEL > 0 -> defined
  DBG_UTIL" and a7c7942f "DBG_UTIL vs
  OSL_DEBUG_LEVEL confusion"

* --enable-dbgutil would no longer prevent -DNDEBUG (i.e., assert became
  ineffective in such builds)

Writer does a lot of work, e.g. spell-checking, word counting etc. in
idle-callbacks. It halts work by checking for AnyInput, and if any input or
paint etc is pending the idle-callbacks stop. With gtk3 rework pending
events don't seem to be available quite right.