fix(node): drop cached node id when the document is updated

[route]

Fixes #86

A Ferrum::Node caches its CDP node id. Chrome invalidates node ids on DOM.documentUpdated and reuses them for the new document, so a node found before a navigation or reload silently acted on another element: focus, click, scroll_into_view, evaluate on the node, or passing it as a script argument.

This also affected Cuprite: its this.isConnected guard resolves the node through the same cached id, so it passed for the wrong element and the command went ahead.

Changes

• Page#document_version is bumped on every DOM.documentUpdated. Node#node_id re-requests its id from its remote object once the version changed. For a node from a previous document this raises NodeNotFoundError, which Cuprite turns into ObsoleteNode, so Capybara reloads and retries.
• The version is recorded only after a successful lookup, so a failed one doesn't leave the stale id looking current.
• NodeNotFoundError raised by Node#node_id now carries its message ("node is not trackable") instead of the class name.

Testing

• New spec: a node found before a reload raises NodeNotFoundError on focus (twice) instead of focusing the new document's element. It fails without the fix.
• Ferrum suite: 633 examples, 0 failures.
• Cuprite suite against this branch: 1863 examples, 1 failure, spec/features/driver_spec.rb:113, which also fails on main. It calls Process#default_user_agent, removed in ref(browser): read browser details over CDP and ask /json/version only for url: #651.