[iamralch]

A first page had no cursor, and a last page returned None. So every caller special-cased the empty page_token on the way in and turned None back into it on the way out. It's the same branch the blank filter needed before #8. AIP-158 reads the empty token as the first page in a request, and as "no more pages" in a response.

let cursor = Cursor::parse(&page_token)?;              // "" -> the empty cursor
query.with_cursor(cursor);                             // empty: no cursor at all
let page = pager.next_page(rows)?;
let next_page_token = page.cursor.encode();            // last page: ""

• Cursor gains a default, the empty cursor (no keys, no position), and is_empty(). Cursor::parse("") (or whitespace) returns it, and encode() of it returns "".
• with_cursor(empty) is no cursor: nothing is AND-ed in. There's also no cursor ordering to compare an explicit order_by against, so a first page can't trip CursorMismatch. A later empty cursor clears an earlier one, just as a second with_cursor replaces the first.
• Page::cursor is a plain Cursor, empty on the last page.
• Pager docs note that over an empty ordering every page is the last, since there's nothing to position by.

Breaking: Page::cursor is Cursor, not Option<Cursor>; check is_empty() for the last page. Cursor::parse("") now returns Ok with the empty cursor instead of a token error.

Tests:

• the empty token and the empty cursor round-trip;
• an empty cursor composes exactly like no cursor, even beside an explicit order_by;
• an empty cursor clears a real one;
• the SQLite and PostgreSQL pager tests now start from the empty token/cursor and stop on an empty one.

The README's keyset section is updated.