ci: make valgrind and coverage require successful check to start

[jodavies]

Avoid burning CPU time and clogging up the runners if the normal checks fail.

Also separate parform from check: it takes a long time

[coveralls]

coverage: 50.536% (-0.02%) from 50.552%
when pulling c235df1 on jodavies:actions-cond-valgrind
into d3f7886 on form-dev:master.

[tueda]

Do we need to make the valgrind-check and coverage jobs require the successful completion of the check-i386 job?

[jodavies]

In principle no, but it runs quickly, and the main motivation was @magv 's iterations on #674 which were failing on the 32bit test in particular.