tests: --ui=record does not work with multicore testing

[matejcik]

AFAICT each runner tries to write its own results, they all write over each other, and the changes are lost. My understanding is that this is essentially random and some subset of changed tests would get written out given the right timing.

In any case, each runner must write their own fixture results, and the xdist controller is the one to merge them into the resulting fixtures.json.

General improvement suggestion: it would be nice if, internally, each test run produced a separate list of ui hashes (then the workers can dump that into fixtures.gwXYZ.json), and this list would then be merged into the main list, keeping or removing pre-existing tests based on --ui-check-missing