fix #17853 (ascii message separator broke json nim dump)

[timotheecour]

• fix Compiler json dump inserts invalid json character at end of output #17853 by making ascii message unit separator (from unit separator #17730) optin (via --msgSep:on); it's only useful for some tooling (eg IDE), and otherwise causes issues such as Compiler json dump inserts invalid json character at end of output #17853 (eg breaking nim dump --dump.format:json)
• add regression test for nim dump --dump.format:json
• remove **keep binary compatible** notice for TGlobalOption
• tests that were modified by adding \31 in unit separator #17730 are now reverted back
• fix a bug in tests/osproc/treadlines.nim which didn't honor testament --nim:/pathto/nim r test

CI failure unrelated (classic flaky #17325)

[timotheecour]

previous code was using "nim", which didn't honor testament --nim:/pathto/nim r test

[Araq]

I prefer to simply fix eg breaking nim dump --dump.format:json) instead. With your philosophy we keep adding more and more of these switches which then are covered by tests (which is good, don't misunderstand me!) -- however, it makes it harder and harder to contribute patches to the Nim compiler for non-experts. Eventually there is always some test failing somewhere for mysterious reasons. Complexity always bites back. And no, putting all these switches into a nice hierarchy does not help much.

[timotheecour]

I prefer to simply fix eg breaking nim dump --dump.format:json) instead

done (but I can't help but remember the similar issues that dot processing caused)
I'm all for fewer options and sane defaults whenever possible, but sometimes, you need them, and when you do it helps that they're kept orthogonal, discoverable etc. In this case, keeping \31 without an off switch is fine i guess.

[timotheecour]

see also #18162 which ended up doing --msgSep:on|off after all