Fix endpoint handling for batched quantile grids

[tandede]

Summary

interpolate_quantiles accepts a tensor of quantile levels matching original_values, allowing each batch element to use a different quantile grid. Endpoint padding was previously decided from the global minimum and maximum across the entire batch.

If one row contained 0 or 1 while another did not, the latter row was left unpadded. Queries below that row's minimum produced a negative gather index, while queries above its maximum were extrapolated instead of using the endpoint value.

This change pads every flattened batch row with 0 and 1 and repeats that row's first and last values. Duplicate endpoints are safe for rows that already contain them.

Test plan

• .venv/bin/python -m pytest -q test/test_utils.py — 7 passed
• .venv/bin/mypy src test — no issues in 24 source files
• .venv/bin/python -m pytest -q — 554 passed, 20 xfailed
• uvx ruff check --select E,F,I,B src/chronos/utils.py test/test_utils.py
• uvx ruff format --check src/chronos/utils.py test/test_utils.py
• git diff --check

[tandede]

Hi maintainers, I wanted to follow up on this PR in case it was overlooked. It fixes endpoint handling for batched quantile grids and is ready for review. Please let me know if any changes or additional validation would be helpful. Thank you!