fix: keep Replicate private endpoint output as one Message

[fallintoplace]

private Replicate endpoints already join prediction.output into one string, but the private endpoint path then iterates that string when building the return value.

so a response like "hello" becomes a list of per-character messages instead of a single Message("hello").

this PR:

• returns one Message for the joined response
• adds a regression test for chunked prediction.output

verification

• ran PYTHONPATH=/Users/hoangvu/Code/OSS/garak /tmp/garak-pr-venv/bin/python -m pytest tests/generators/test_replicate.py
• test passes with this change
• the same test fails against the old behavior because _call_model() returns one message per character

I didn't run a live private Replicate endpoint check here, and I didn't run the full tests/ suite in this local environment.

[aishwaryap]

This PR appears to fix a legitimate issue but does not yet meet the bar for acceptance. Unit tests alone are not sufficient for this fix. We require evidence that this fix has been tested against an actual Replicate endpoint. If you are not comfortable placing such evidence in a public PR, please email it to garak@nvidia.com with the subject "Evidence for PR #1836".