Fix assert on New Layer From Visible on invisible active layer

Before this commit, Krita would assert on New Layer from Visible
when used on invisible layer. This commit fixes this.

Explanation:
Buckle up, it's going to be a ride :)

In 2016 there was a problem that invisible layers would show up
when they were merged, see bug 359707.
https://bugs.kde.org/show_bug.cgi?id=359707
As a solution, "just remove the layers separately instead of
merging" strategy was used.

In 2018 it turned out that in case when putAfter (the layer
Krita puts the merged result above of) is invisible, it gets
removed, so then it's null and Krita tries to merge other layers
in instead of putAfter, see bug 395903.
https://bugs.kde.org/show_bug.cgi?id=395903
As a solution, if putAfter was invisible, one of the visible
nodes-to-be-merged were assigned to putAfter to replace the
original putAfter node, to make sure that when original putAfter
is removed, some other layer is doing its job.
(The fact that it's a soon-to-be-merged layer is not a problem).

In 2020 it turned out that there is quite a lot of actions
that are possible in locked groups. It required some more changes
to the code. See bug 406697.
https://bugs.kde.org/show_bug.cgi?id=406697
A solution included a change to the relevant code such that
invisible putAfter would be unconditionally replaced with
one of the nodes from the to-be-merged list, and then
if putAfter had no parent, Krita would encounter an assert.

Problem with this behaviour:
In case of newLayerFromVisible(), mergedNodes list would only
contain a root node (which has no parent). If the putAfter is
invisible, then it's replaced with "one of" mergedNodes, which
means the root node. Which means it's going to assert.

However it's not always necessary to replace putAfter with one
of the to-be-merged nodes. It's only needed when it's going to
be removed. Hence moving the replacing into the if.
Because root is not invisible, the list of invisible nodes
will be empty and nothing will be replaced and nothing will
be removed.

NOTE: I also added a new condition: '&& cleanupNodes'.
It's not stricly necessary to cover all usecases of
mergeMultipleNodesImpl, because the only usecase where
cleanupNodes is false is the newLayerFromVisible case
(which will fail the first condition).
However it makes it more future-proof since the only case
those nodes needs to be removed is when cleanupNodes is true
(otherwise removing nodes would be a dataloss bug).

NOTE2: It's possible that putAfter replacement needs one more
check (to see if it is in fact in the invisibleNodes list).
I couldn't find a situation  where it was needed, though,
so I left it as it is.

CCBUG:359707
CCBUG:395903
CCBUG:406697

BUG:428683


(cherry picked from commit c1ab1ee9)