Make sure that transient errors are returned correctly

Following PR #806 which removed various grids from nad/, we now have
to download the BETA2007 grid before that DHDN_ETRS89 test can pass
successfully.

For reasons unknown to anyone but the geniuses from Redmond: In
MSWindows, the wildcard pattern "*.gie" expands to all files having
an extension *starting* with ".gie".

In other words "dir *.gie" will list files with extension ".gie"
*and* with extension ".gie_failing".

This means that running "gie test\gigs\*.gie" will run all gigs
tests, even the ones known to fail.

So rather than getting a quick check for regressions, one gets
a lot of noise and no easy way to be sure whether any regressions
have appeared.

To work around this oddity, we rename all files ending in
".gie_failing" to end in ".gie.failing" - and while at it,
adding all the non-failing gigs tests to the AppVeyor test suite.

This extends the regression test suite, run at the end of each
AppVeyor integration, to over 2500 individual tests.

Test coverage, however, is still hovering just below 75%, so
there's still plenty of room and reason for additional tests.

Copy ellipsoid definition for proj=cart directly into the
proj_create call, rather than calling pj_inherit_ellipsoid_def
afterwards.

Previously, the ellipsoid definition was left out from the call.

pj_init_ctx would then pick up WGS84 from proj_def.dat, and the
init would succeed (and the possibly wrong ellipsoid definition
would later on be overwritten with the correct values by
pj_inherit_ellipsoid_def.

But if PROJ_LIB was not set or proj_def.dat was inaccessible for
other reasons, things went wrong.

Reset error type PJD_ERR_MAJOR_AXIS_NOT_GIVEN for operations
that do not need an ellipsoid.