Add cecil-data-analysis-xql skill

[jayendra13]

Summary

Adds a new skill, skills/cecil-data-analysis-xql/, that answers earth-observation analysis questions against Cecil datasets using xarray-sql (xql) as the SQL layer.

The skill owns the full loop:

1. Pick the right dataset from the Cecil catalog (prefers existing subscriptions; gates new ones on explicit confirmation).
2. Load it via client.load_xarray(subscription_id).
3. Register the dataset and every variable's reference_table on an xarray_sql.XarrayContext.
4. Run the query and present SQL → result table → plain-English interpretation as one compact block.

Why xql: SQL beats pandas chains for joining categorical reference tables (no integer codes leaking into answers), windows like ROW_NUMBER OVER / LAG OVER make dominant-class and pixel-level change-detection queries one-liners, and DataFusion streams over the dask-backed xarray Dataset the SDK already returns.

What's in the skill

• SKILL.md — instructions, golden-rule output format, worked example
• references/ — Cecil dataset catalog pointers, condensed SDK reference, xql patterns + gotchas, and the canonical demo script
• scripts/ — list_subscriptions.py, inspect_dataset.py, run_analysis.py (load → register → run SQL → save result.csv + result.md + result.png)

Cross-references

Per CONTRIBUTING:

• Prerequisite: subscribe-and-load — sets up the subscription load_xarray expects.
• Hand-written counterpart: land-cover-baseline-and-change — same outputs without a SQL layer.

Skill checklist

• Directory name (cecil-data-analysis-xql) is kebab-case and matches the name: in frontmatter.
• Frontmatter contains name, description, license: MIT.
• Body covers prerequisites, steps, constraints, references.
• Cross-references to sibling skills use relative paths.

Test plan

• Smoke test: list_subscriptions.py — returns 2 subscriptions on Land Cover 9-Class, output matches the documented format.
• Smoke test: inspect_dataset.py against the Land Cover 9-Class dataset — returns 9-class reference table cleanly.
• End-to-end: dominant-land-cover-class worked example through run_analysis.py — runs the windowed SQL, joins the reference table, writes result.csv / result.md / result.png, prints the Markdown summary.

End-to-end output:

|   year | dominant_class   |      px |
|-------:|:-----------------|--------:|
|   2020 | Trees            | 9167795 |
|   2023 | Crops            | 8876719 |

The end-to-end test caught one missing transitive dependency (tabulate, needed by pandas.DataFrame.to_markdown()); fix included in the second commit on this branch.