The old code relied on the order of initializing global static variables across
compilation units, which failed for MinGW cross compilation.  Cleaned up.

...as soffice.exe creates the proper environment (PATH) for soffice.bin to be
able to load URE libs like sal3.dll, so soffice.exe itself should neither be
able to load nor depend on sal3.dll (and does not depend on anything exported by
it, anyway).

It is unclear to me how this ever worked (which it presumably did?) with any
Windows build, either native MSVC or cross-compiled MinGW.  (Maybe MSVC builds
do not record a dependency from soffice.exe to sal3.dll, noting that the former
does not import anything from the latter?)

Or should it be --without-helppack-integration that affects it? Or
both? Why do we need both these switches?  And what does that
irritating "integration" thing mean again? Does
--without-helppack-integration mean "yes, do build helppacks, but
don't integrate them"?

libxslt in Fedora mingw packages depends on it.

Writer's default is 0, but Word's default is 26 (15 in 1/8 pt)

Otherwise code using the lcms2 we build won't get the stdcall
decorations and linking will fail.

Nah, will use the patch by Lubos instead.

This reverts commit 85f07455.

The idea is to use libo_MINGW_CHECK_DLL for libs that must be available
(typically that would be the "main" library, e.g., libxml2 or libcurl)
and libo_MINGW_TRY_DLL for possible dependencies (that may not be the
same on different systems). All further references to the dlls are
exclusively through the configured variables or defines set from these
variables (e.g., instead of hardcoding libxml2-2.dll, use
$(MINGW_LIBXML2_DLL)).

The macros are documented in m4/mingw.m4 .

Files that must be changed when adding a new dll:
* configure.in
* config_host.mk.in
* external/mingw-dlls/makefile.mk
* scp2/source/ooo/makefile.mk
* scp2/source/ooo/mingw_dlls.scp

Use GG era abbreviation in one format instead of full GGG.

Hopefully this will allow pasting to external applications with whole
column/row selected while not affecting calc itself.

These two tests are not well designed and need 5 and 10 minute just to
check that we really return every row. This is not a good trade-off
between test coverage and execution time.

Enable them again with a better concept.