Fixes #2752.

Summary

examples/14-transfer-learning.ipynb fails on the air dataset with ValueError: ... NaN ... errors during model training. The raw carrier_passengers.csv contains gaps that survive longest_contiguous_slice() (e.g. zero-rows the slice helper does not flag as missing), so the resulting series carry NaNs into model training. As @rafmacalaba reported in #2752, the cleanest workaround is to call fill_missing_values(...) after the slice; @dennisbader acknowledged the notebook needs refactoring (it's one of the few examples that's not run in CI because it's slow).

This PR applies that fix in the smallest possible patch:

• Import fill_missing_values from darts.utils.missing_values in the imports cell.
• Inside load_air(), call fill_missing_values(series, fill="auto") immediately after longest_contiguous_slice(). An inline comment cites the issue so a future reader doesn't have to dig.

Other Information

• No code in the darts/ package is touched. Only examples/14-transfer-learning.ipynb and CHANGELOG.md.
• CHANGELOG entry under Unreleased > Fixed.
• Maintainer-stated that the notebook isn't part of CI, so adding a unit test wouldn't fit existing conventions. The notebook change is itself the regression artifact: the diff is what unblocks users running the example end-to-end.

Reproduce BEFORE/AFTER yourself (copy-paste)

# --- one-time setup ---
git clone https://github.com/unit8co/darts.git /tmp/darts-2752 && cd /tmp/darts-2752
uv sync --group dev

# --- inspect the difference (no need to run the full notebook) ---

# BEFORE (origin/master): load_air does NOT call fill_missing_values
git checkout origin/master
grep -n "fill_missing_values" examples/14-transfer-learning.ipynb || echo "NOT PRESENT"
# Expected: "NOT PRESENT" — the helper is neither imported nor called.

# AFTER (this PR): load_air now imports + calls fill_missing_values
git fetch https://github.com/jbbqqf/darts.git fix/2752-transfer-learning-nan
git checkout FETCH_HEAD
grep -n "fill_missing_values" examples/14-transfer-learning.ipynb
# Expected: 2 hits — one in the imports cell, one inside load_air()
#           right after longest_contiguous_slice().

If you want to verify functionally (slow — downloads the dataset and runs the local-models cells):

# AFTER, end-to-end smoke
uv run jupyter nbconvert --to notebook --execute examples/14-transfer-learning.ipynb \
    --output /tmp/14-executed.ipynb --ExecutePreprocessor.timeout=600
# Expected: no NaN-related ValueError; eval_local_model finishes with finite SMAPE values.

What I ran locally

• Validated the notebook is well-formed JSON and contains 76 cells (unchanged count).
• Confirmed fill_missing_values import + call are present in the patched notebook.
• Confirmed the diff is limited to the imports cell and the load_air() cell (plus the CHANGELOG).
• Did not execute the full notebook locally because it requires downloading the M3 / M4 / carrier_passengers datasets (the same reason maintainers don't run it in CI per [BUG] examples/14-transfer-learning.ipynb error on air_train because of missing values #2752).

Edge cases tested

Risk / blast radius

• Notebook-only change. No runtime code in darts/ is modified.
• fill_missing_values(..., fill="auto") is a public, documented helper used elsewhere in the codebase, so the new dependency is benign.

PR drafted with assistance from Claude Code (Anthropic). The change was reviewed manually against unit8co/darts@master, the discussion in #2752, and the proposed fix from the original reporter. Reviewer can paste the BEFORE/AFTER block above to verify.