fix(cli): scrape progress, working Back, multi-URL Enter; scrape folder as dataset input

[ianktoo]

Summary

Three fixes for the interactive scrape and wizard flows, plus a new input option.

Fixed

No feedback while scraping several URLs

• Before: Scrape pages (no AI) showed one spinner until every page was done, and dataforge scrape printed nothing until the end. Pages on the same site are fetched one at a time to respect its rate limit, so a list could look stalled for a long time.
• Now: scrape_urls takes a progress callback (called before and after each URL). Both entry points show a live bar naming the page being fetched (Fetching 2/5) and print a line for each page as it finishes. With --json, the lines go to stderr, and only when stderr is a terminal, so stdout stays clean JSON.

Back did not go back

• Cause: questionary gives a Choice with value=None its title as the value. Back in ask_input_method, ask_discovery_scope and ask_adjust_settings_target returned strings like "(b) Back to menu" instead of None.
• Effect: the wizard re-showed the same step, and Back on the scope question saved "(b) Back" as discovery_scope.
• Now: Back choices use a sentinel that is mapped to None. Backing out of a URL box returns to the method list, Back on the method list returns to the main menu, and "back" on the first wizard step goes home instead of repeating it.

Multiple URLs box looked frozen

• Cause: the prompt said "blank line to finish", but questionary's multiline box only finishes on Alt+Enter, so Enter on a blank line just added another line.
• Now: Enter on an empty line finishes. An empty box shows a validation message instead of submitting nothing, and the number of URLs added is printed.

Added

Pages I scraped earlier (in Build an AI training dataset)

• Choose a folder written by the no-AI scrape. The default is the newest folder in ./scrape/.
• scrape.load_scrape_dir reads pages.jsonl, or the page_NNN.md files when only Markdown was saved.
• scrape.import_into_session writes the raw files under sessions/<id>/raw and adds DiscoveredURL and ScrapedPage rows, so the later stages and resume work unchanged.
• The run starts at processing. The robots.txt and skip-known questions are skipped because nothing is fetched.
• After a menu scrape, a hint points to this option.

Tests

• New tests/test_wizard_navigation.py: key-press tests for Back in each affected menu, the multi-URL box (finish on an empty line, empty Enter re-asks), backing out of a URL box, Back on the first wizard step, and choosing a scrape folder without being asked the scope.
• tests/test_quick_scrape.py: the order of progress callbacks, loading a scrape folder from JSONL and from Markdown only, and imported pages going through ProcessorAgent.
• The full suite and ruff pass locally.

Docs

README (scrape section), CHANGELOG (Unreleased) and TECHNICAL.tex (discovery section) updated.

Not in this PR

dataforge run recipes and the MCP start_run tool cannot take a scrape folder yet. A source.scrape_dir recipe field would cover both.