fix(examples): migrate anomaly-detection example to TimesFM 2.5

[ihubanov]

Companion to #465, for the other example still on the removed v1 API. Independent of it; either can merge first.

detect_anomalies.py fails on the first model load:

AttributeError: module 'timesfm' has no attribute 'TimesFmHparams'

Changes

• Load via TimesFM_2p5_200M_torch.from_pretrained + model.compile(ForecastConfig(...)), using the same config as run_forecast.py in the global-temperature example.
• Pass horizon explicitly to forecast() and drop freq=[0].

The quantile column mapping in this file was already correct (IDX_Q10, IDX_Q20, IDX_Q80, IDX_Q90 = 1, 2, 8, 9), so only the loading path changes. The diff is 20 lines.

Behaviour change in the regenerated artifacts

2.5's continuous quantile head gives noticeably tighter intervals than the checkpoint the committed artifacts came from:

80% PI width	month 1	month 12
before	0.227	0.494
after	0.216	0.329

One more month therefore falls outside the band, so the forecast CRITICAL count moves 4 -> 5. Worth flagging rather than burying, though it does not weaken the example:

• All three injected anomalies are still detected (2025-04, 2025-09, 2025-12).
• Phase 1 is pure numpy detrend + Z-score and is unaffected: the acceptance criterion in SKILL.md, Sep 2023 CRITICAL, still holds at z=+3.03.
• The extra detection is 2025-05, a noise month that now sits just outside the narrower 80% band. 2025-11 was already a false positive before this change.

If you would rather keep the old detection counts, the honest lever is the injected offsets in build_synthetic_future() rather than loosening the bands. I left them alone, since tuning the fixture to preserve a number seemed worse than reporting the difference.

Verification

Ran end to end on CPU (Python 3.13, torch 2.13.0+cpu). Regenerated anomaly_detection.json and anomaly_detection.png.

Refs #423

[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.

[ihubanov]

@googlebot I signed it!