fix: np.float128 platform crash on macOS/Windows & missing f-string in error message

[prince-shakyaa]

Fix: np.float128 platform crash & missing f-string in error message

Closes #657
Closes #658

---

What this PR does

This PR fixes two backend bugs found during local development on macOS Apple Silicon (ARM64).

---

Fix 1 : Guard np.float128 for platform compatibility (Closes #657)

File: gemma/gm/utils/_dtype_params.py (via kauldron dependency)

np.float128 is not available on macOS Apple Silicon or Windows — it only
exists on Linux x86-64 as a long double alias. Because kauldron/ktyping/dtypes.py
unconditionally accesses np.float128 at module import time, every test
that transitively imports from kauldron.ktyping fails immediately during
pytest collection with:

AttributeError: module 'numpy' has no attribute 'float128'. Did you mean: 'float16'?

This caused 162 test collection errors on Apple Silicon in the CI-equivalent
local run. The fix guards the attribute access with hasattr:

# Before:
float128 = NpDType(np.float128)

# After:
float128 = NpDType(np.float128) if hasattr(np, 'float128') else NpDType(np.float64)

Note: The root cause is in the kauldron dependency. This fix is being
upstreamed there. On the gemma side, the CI should be pinned to a kauldron
version that includes the fix once merged.

---

Fix 2 : Add missing f-prefix in _normalize_token() error message (Closes #658)

File: gemma/gm/text/_sampler.py, line 579

The error raised when a multi-token stop_token or forbidden_token string is
supplied was silently printing {token!r} as a literal string rather than
interpolating the actual bad value, because the f prefix was missing.

# Before (plain string - unhelpful):
raise ValueError(
    'Invalid token: {token!r}. `stop_token`s and `forbidden_token`s must'
    ' map to single token ids in the vocab.'
)

# After (f-string - shows the actual token):
raise ValueError(
    f'Invalid token: {token!r}. `stop_token`s and `forbidden_token`s must'
    ' map to single token ids in the vocab.'
)

---

Testing

• Existing tests pass locally (after the kauldron fix is applied).
• The _normalize_token error path manually verified to produce the correct interpolated message.

No new tests are added — Fix 1 is a dependency-level change and Fix 2 is a
one-character typo correction.

---

Checklist

• Code follows the Google style guide (80-char lines, 2-space indent, pyink formatted).
• All existing passing tests still pass.
• PR is linked to the relevant issues (Bug: Test suite fails on macOS Apple Silicon & Windows — does not exist #657, Bug: Missing f-prefix in error message in _normalize_token() makes stop_tokens debugging impossible #658).
• No new dependencies introduced.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[prince-shakyaa]

I signed it!

[prince-shakyaa]

Note: The recent force-push on this branch was just an empty amend to re-trigger the Google CLA bot checks after signing the agreement, as the webhook seemed to be stuck. The code changes remain exactly the same!

[prince-shakyaa]

While testing locally, I noticed the ValueError raised by _normalize_token was literally printing {token} instead of the actual token variable. This PR simply adds the missing f prefix to correctly interpolate the token string!

Let me know if you need me to make any changes.
Thank You.