Fix stale flavortag on shared wavefunctions in HelasDiagram.check_flavor

[oliviermattelaer]

HelasDiagram.check_flavor only cleared the flavortag attribute on the wavefunctions introduced by the diagram (self['wavefunctions']). In a HELAS-optimised matrix element a diagram also reuses wavefunctions (and their sub-trees) introduced by earlier diagrams, which appear here only as mothers. Those shared mothers kept a stale flavortag computed for a different flavor, and propagate_flavor_tag silently reused it instead of recomputing it. This made check_flavor order-dependent: a diagram whose internal lines are all reused (e.g. the t-channel gluon diagrams of q q > q q g and q q~ > q q~ g where the external gluon is radiated off a final quark) was wrongly rejected for a valid same-flavor assignment when checked after another flavor.

Clear the flavortag over the whole ancestor closure of every wavefunction and amplitude instead. External leaves are re-tagged as before; cleared internal shared mothers are recomputed by propagate_flavor_tag. This is the same bug class previously patched at the check_flavor_for_all_diagrams level; the fix here makes the lower-level HelasDiagram.check_flavor correct on its own.

Add regression tests (TestDiagramFlavorCheck, TestDiagramFlavorCheckQQG, TestDiagramFlavorCheckQQbarG) covering d d~ > d d~, q q > q q g and q q~ > q q~ g; the latter two fail without this fix.

[oliviermattelaer]

@theoheimel do you want to do some test on this or we can merge it?