Fix missing f-prefix in _normalize_token error message

[HumphreySun98]

Summary

_normalize_token() in gemma/gm/text/_sampler.py raises a ValueError
when a stop_token / forbidden_token string does not map to a single
token id. The message string was missing its f prefix, so {token!r}
was printed literally instead of interpolating the offending token,
making these misconfigurations hard to debug.

Fixes #658.

Change

-        'Invalid token: {token!r}. `stop_token`s and `forbidden_token`s must'
+        f'Invalid token: {token!r}. `stop_token`s and `forbidden_token`s must'

The second (implicitly concatenated) line has no placeholders, so it is
left unchanged.

Testing

Added test_normalize_token_error_message_includes_token, which drives the
error path with a fake tokenizer and asserts the message names the token.
Verified it is a real regression guard — it fails on the pre-fix code
and passes with the fix.

$ python -m pytest gemma/gm/text/_sampler_test.py -q
5 passed

[wbizmo]

Cloned it and diffed against main.

_sampler.py gets the missing f prefix, _sampler_test.py has a test for it. Ran the string through both versions myself, old code literally prints {token!r} instead of the token, new one shows you 'multi_token'. Test regex catches the difference too, fails on the old string, passes on the new one.

Approving.

[HumphreySun98]

Cloned it and diffed against main.

_sampler.py gets the missing f prefix, _sampler_test.py has a test for it. Ran the string through both versions myself, old code literally prints {token!r} instead of the token, new one shows you 'multi_token'. Test regex catches the difference too, fails on the old string, passes on the new one.

Approving.

Thank you so much for your review!

[HumphreySun98]

@Qwlouse gentle follow-up when you have a moment 🙂 Small, behavior-preserving fix for #658 — _normalize_token was missing its f-prefix, so the error printed the literal {token!r} instead of the offending token. All checks green, cleanly mergeable, and covered by a regression test. Happy to rebase or adjust anything. Thanks for taking a look!