fix: preserve single-knot KTR predictions with uncertainty - #896

Description

Fixes #816.

KTR prediction with level_segments=0 accesses a nonexistent second level knot when include_error=True. Only extrapolate level knots when at least two knots exist. Single-knot predictions use the existing constant interpolation kernel; observation noise remains enabled.

Regression tests cover in-sample and future dates, one and three posterior samples, and predictions with and without observation noise. They also check that prediction leaves posterior levels unchanged.

Type of change

• Bug fix

How Has This Been Tested?

• Before the fix: four regression cases fail with the reported IndexError; four pass.
• After the fix: python -m pytest tests/orbit/utils/test_knots.py tests/orbit/models/test_ktr.py -k 'single_level_knot or test_knots' -q — 29 passed.
• black --check orbit/template/ktr.py tests/orbit/models/test_ktr.py — passed.
• git diff --check — passed.
• Broader KTR fitting tests stop because this Windows environment has no CmdStan installation. Full-suite success and the required coverage threshold are not verified, and still need verification before merge. The code change is ready for review with these validation limits.

No dependency changes. Prepared with Codex assistance.