check.milter: fail_open does not cover a milter session-dial failure

## Summary

`check.milter`'s `fail_open` directive is documented to let a message through when the milter can't be consulted, but it silently doesn't cover one specific failure mode: the milter being completely unreachable when maddy tries to establish a session.

## Root cause

`CheckStateForMsg` (`internal/check/milter/milter.go`) dials the milter backend via `c.cl.Session()` and returns any error unconditionally. This happens *before* a `*state` exists, so the failure never reaches `ioError()` — the only place `fail_open` is actually checked. Result: with `fail_open` set, a message is still hard-rejected (a real SMTP-level error handed back to the original sender) if the milter is simply down — exactly the case `fail_open` exists to guard against.

## Reproduction

Configure `check.milter` with `fail_open: true` pointing at a milter address that's genuinely unreachable (e.g. nothing listening on that port), then send a message through. Expected: message accepted unchecked, per `fail_open`. Actual: a hard SMTP-level rejection.

## Fix

I've prepared and tested a fix — handles the dial failure the same way `ioError()` already handles a later I/O failure — and opened it as #865, with a regression test that fails against unpatched code and passes with the fix.

Apologies for the PR arriving before this issue — I realize CONTRIBUTING.md's process is to raise the issue first. Filing this now so there's a proper place to discuss the bug alongside the PR.

Lisez internal/check/milter/milter.go, en particulier CheckStateForMsg et la gestion existante de ioError. Examinez le test de régression mentionné dans PR #865 et vérifiez qu’un milter inaccessible avec fail_open activé accepte le message sans le vérifier, tandis que le test réussit.

Rédigé par le modèle d'indexation à partir du texte de l'issue.