Add window bounds to list_windows and get_active_window

list_windows returns {title, pid, handle, className} and get_active_window returns {title, pid, handle}. Neither returns window geometry.

Impact

Without bounds an agent cannot:

Take a cropped, per-window screenshot — the single biggest token lever in computer-use work (frontier harnesses crop to the relevant window rather than sending 1920x1080 every step).
Convert window-relative coordinates to screen coordinates.
Tell whether a window is offscreen, minimized, or occluded before clicking into it.
Verify that a move_window / resize_window call actually took effect.

get_screen_info already returns full monitor bounds, so the response shape exists — it just stops at the monitor level.

Add bounds: {x, y, width, height} (via GetWindowRect) plus isMinimized / isVisible to both list_windows and get_active_window.

This is also the prerequisite for a UI Automation perception layer (per-window tree scoping), which is the follow-up sprint.