Move the grid download loop into an importable scrape package

[daniel-widrick]

PR 4 of #15. Stacked on #19.

Changes

• New scrape package with Fetch(ctx, prefs, opts) (*guide.TVGuide, error). It downloads every six-hour slot for a lineup and assembles the guide: Channels deduplicated by station in first-seen order, Lineup with every position, Programs deduplicated by station and time, and Source. No enrichment, no file I/O, no env reads.
• Options covers days, inter-slot delay, an injectable GridFetcher (which *web.Client satisfies), a clock, a per-slot Progress callback, and a logger. The zero value is production configuration.
• Context cancellation is honored before each slot, inside the request, and during the inter-slot delay.
• main.runScrape calls scrape.Fetch, maps progress onto the setup status page, and cancels the context when the active lineup changes mid-download (polled once per second, with a live re-check after the fetch).
• README gains a "Using as a Library" section; CLAUDE.md and the project structure are updated.

Behavior changes to be aware of

• When every grid slot fails, the scrape now returns scrape.ErrNoData instead of writing an empty guide. The previous guide stays live and the existing 15-minute retry applies.
• XMLTV channel order is now first-seen order instead of Go map iteration order, so it is stable between runs.

Tests

• scrape: slot alignment and count; full assembly with a station at two numbers; distinct events across slots kept while repeats are deduplicated; a failed slot skipped and reported with progress counts monotonic; ErrNoData when all fail; cancellation during the delay, during a request, and before the first request; option defaults. Runs clean under -race.
• main: runScrape end to end with a fake fetcher (lineup assembled, proxy rewrite applied, files written, progress stages reported); a lineup change is reported as errScrapeSourceChanged and the persister never runs; total failure surfaces ErrNoData and the persister never runs.