CI: Fix the relocated installation test and instrument the Windows failure

[petrasovaa]

Experimental branch for iterating on the CI failures of OSGeo#7733. Not meant for OSGeo/grass as it is.

Linux and macOS: the new step could not import grass at all

The step Run pytest for loading C libraries into the process failed with exit code 4:

ImportError while loading conftest '.../python/grass/script/tests/conftest.py'.
python/grass/script/__init__.py:4: in <module>
E   ModuleNotFoundError: No module named 'grass'

PYTHONPATH was set to ${HOME}/install-relocated/etc/python, but Autotools installs into the prefix plus a version directory, so the real path is ${HOME}/install-relocated/grass86/etc/python. With a non-existent entry there, grass is not importable and pytest falls back to importing the repository copy of grass/script/__init__.py by file path, which then does not find grass.exceptions. The path is now taken from grass --config python_path before the move (the command does not work from the moved installation) and rebased onto the new prefix.

Windows: instrumented, not fixed yet

test_grass_lib_usable_with_load_libs fails in the main pytest step (which is why the new step never runs there). The subprocess exits with 1 and empty stdout and stderr - no Python traceback, so this looks like a fatal error in the C library, whose message is lost. Timing suggests it got as far as the grass.lib.raster calls.

The sibling test with a custom environment passes even though its session variables exist only in a dict that the C library cannot read, which means G_gisinit there succeeds against some other session. That is consistent with the limitation recorded in lib/gis/tests/lib_gis_env_test.py (the C runtime keeps its own copy of the environment on Windows), but the logs do not prove it.

So this round only makes the failure legible:

• the failure message includes both output streams,
• the first test writes the session the C library resolved (GISDBASE, LOCATION_NAME, MAPSET) to stderr, flushed, before the raster is opened,
• the custom environment test asserts that this session is the one init() created instead of accepting any working session.

Depending on what the Windows run reports, the next step is either skipping that test on Windows with the cached-environment reason, or pushing GISRC, GISBASE and GIS_LOCK into the C runtime with libgis.G_putenv at the end of init() when load_libs is set.

The analysis of the logs and these changes were done with Claude Code.