Fix browser tools hanging for minutes on dialogs and stuck pages

[stippi]

Problem

The browser tools sometimes hung for minutes and then failed with "Request timed out". The engine is chromiumoxide (CDP), which lacks protections Playwright provides out of the box. A local test reproduced it:

• A click that opened an alert() did not return within 120 s.
• After that, every CDP command waited for its timeout. A screenshot took about 16 minutes before "Request timed out".
• chromiumoxide never answered the dialog, so it stayed open and the next tool call hung the same way.

chromiumoxide's own 30 s command timeout does not hold while the renderer is stuck.

Changes

• JavaScript dialogs are answered as they open. alert and beforeunload are accepted. confirm and prompt are dismissed unless browser_act gets accept_dialogs: true for that call. Each answered dialog is reported to the model (Note: a confirm dialog "…" was dismissed.).
• Timeouts: every BrowserSession verb is bounded: 15 s for an interaction, 30 s for a navigation (BrowserTimeouts). A limit that runs out is a typed BrowserTimeout. wait_for and settle are bounded as a whole.
• Capture: when the page cannot be read, capturing the result stops and reports that the page is not responding, instead of also waiting out the screenshot.
• Slow loads: a navigation whose load event never fires (a slow iframe or tracker) is no longer a failure. The page is shown as far as it got, with a note. The login handoff tolerates this too, so a slow portal page cannot fail a login the user already completed.
• Typing: characters outside the US layout (ü, ß, €) used to fail with "Key not found". They are now inserted with Input.insertText.
• Smaller fixes: navigate() dropped its second wait_for_navigation, which has no timeout in chromiumoxide. LaunchedBrowser::close kills the browser if it does not exit within 10 s.

Notes for reviewers

• I chose dismiss as the default for confirm on purpose. Accepting could trigger an outward action, such as submitting a form, whose warning the model never saw.
• WebClient::fetch (the web_fetch tool) has the same unbounded wait_for_navigation after new_page. It is not part of this PR.
• test_web_search needs network access to DuckDuckGo. It failed locally because the network was unavailable, not because of this change.

Testing

• New engine tests in crates/web/src/tests.rs:
  • alert and confirm, including the opt-in
  • typing umlauts
  • a navigation that never loads
  • a page stuck in a script loop: every verb must give up within its limit
• New tool tests in crates/code_assistant_core/src/tools/impls/browser.rs:
  • dialogs reported through browser_act
  • a slow load shown with a note
  • an act on a hung page reported in under 4.5 s
• cargo test --package web and cargo test --package code_assistant_core pass. cargo clippy --all-targets --all-features -- -D warnings and cargo fmt --check are clean.