Fix dry run with messages

[aurelien-reeves]

Fixes #1496 and #1488
I've just applied the suggestion from @cbochs in #1496, it seems to work as expected.

It did not get well previously actually :(

[aurelien-reeves]

So, it looks like the order of the filters is very important 😅

[mattwynne]

I think we should not do this. We only want to publish reports for what's on the main branch.

I understand why you did this, and I think what we need instead is a unit or acceptance test that reproduces the problem that surfaces here when you run cucumber --publish.

[aurelien-reeves]

I am stuck trying to add an acceptance test which reproduce the issue

This is because the tested code is also the testing code. Because of that we cannot have acceptance tests based on the execution order of the filters in the runtime because it is actually executed a first time before executing the test itself.

At the moment I have not been able yet to write a spec either.

[mattwynne]

We decided that adding a test for this is too hard right now, as:

1. It's impossible to use an acceptance test, as the Runtime is already loaded in the acceptance tests
2. The formatter unit tests currently duplicate code from the runtime in a helper (see TODO added in 2667c93)