refactor(download): split the download.py god module into a package

[dylanuys]

Splits the 1889-line dataset/download.py into a cohesive download/ package. Pure code-move — no logic changes.

Structure

constants (errors, extensions, archive-type predicates) ← fetch (HTTP/HF downloading) / listing (remote enumeration + URL resolution) / extract (parquet/tar/zip/raw decode) / cache_io (persistent-cache load) ← core (download_and_extract orchestration, incl. gasstation).

download/__init__.py re-exports the public API, so every existing from ...dataset.download import X keeps working unchanged (verified against all call sites: iterator.py, gasstation_utils.py, dataset/__init__.py, and the parquet test).

How it was done (low-risk)

• Function bodies carved by exact line range (sed), so they are byte-identical to the original — no hand-editing of logic.
• Cross-module imports were driven by ruff: F821 (undefined-name) told me exactly which sibling each module needed; F401 --fix trimmed the shared header down per module.
• Dependency graph is a clean DAG (no import cycles): constants ← {fetch, listing, extract, cache_io} ← core.

Verification

• Def/class inventory identical: 31 defs in, 31 out, diff empty — nothing dropped or duplicated.
• Full import graph loads (no circular imports); public API + external late-imports resolve.
• ruff check src tests clean.
• test_parquet_processing.py: 12/12 pass. Full unit suite: 46 pass, same 5 pre-existing test_config_loading data-drift failures (unrelated).

Note

The _process_*/list_* helper names (incl. the leading-underscore ones) are preserved verbatim to avoid touching call sites; a follow-up could tidy naming/visibility now that they're grouped, but that would no longer be a pure move.