fix(tests): make --skipall pass match the real run

[mkoura]

The initial --skipall registration pass produced allure results that disagree with the real testrun, so the 'Skipped: collected, not run' marker count did not equal the number of collected tests:

• pytest evaluates skipif markers before plain skip markers, so tests with their own skipif (SKIPIF_BUILD_EST_1199, SKIPIF_WRONG_ERA, ...) were registered under that marker's reason and then produced a second, identical result file in the real run. Use a prepended skipif(True) marker instead - it is evaluated before all of the test's own markers, so every collected test is registered with the 'collected, not run' reason.
• The early return in pytest_collection_modifyitems skipped the dynamic marker additions (e.g. the dbsync marker), so a MARKEXPR relying on them selected different test sets in the skipall pass and the real run. Apply the skipall markers after the dynamic markers instead of returning early.

Also remove the dead xdist guard: 'slaveinput' was renamed to 'workerinput' in xdist 2.0, so the guard never fires - and it must not, since with xdist only the worker nodes collect tests and MARKEXPR filtering relies on the markers added by this hook. 'Fixing' the attribute name would have silently broken marker-based selection.

Update the count_test_results.py docstring that described the old registration behavior (pre-fix result directories are still handled correctly by the newest-wins rule).

Verified on a standalone reproduction: with the old behavior, 4 of 7 tests (class-level skipif, decorated skipif, param-level skipif, plain skip) were registered under their own reasons and '-m db' selected nothing during the skipall pass; with the new behavior all 7 register as 'collected, not run' and marker expression selection matches the real run exactly.