fix(render): clear display and reset graphs on new telemetry session

[ololonly]

Closes #30.

Problem

The render loop relied solely on Ratatui's diff rendering and only cleared the terminal twice (entering Debug from loading, and on the first telemetry frame). Any display pixel corrupted by a transient SPI glitch (EMI / ground noise, most likely when the machine switches its relay/heater) was therefore never repaired — Ratatui's previous-buffer still considered the cell correct — so artifacts accumulated until reboot.

Approach

Introduce a session model: when telemetry resumes after the machine has been offline (no UART frame for MACHINE_OFFLINE_TIMEOUT = 30s) or on the first frame, treat it as a new session and request a full terminal.clear(), which heals accumulated artifacts at the natural boundary (machine turned back on).

The clear is now driven by a needs_terminal_clear flag on GlobalAppState that the state machine sets and the render loop drains, replacing the scattered terminal.clear() calls in both the device and simulator loops (also clears on Debug toggle).

MachineState::reset_session() clears the graph buffers (so the chart restarts instead of jumping down from stale data) and drops last_frame/shot_started_at, so the resuming frame is treated as a fresh start and cannot emit phantom mode/shot events (which would otherwise reach MQTT and miscount cups in Home Assistant).

Bundled UX fixes

• Wake-only first press: when the backlight has timed out, the first button press only wakes the backlight and no longer also switches screens.
• Offline waiting screen: telemetry staleness now drives the UI via a new GlobalAppState::machine_online() helper used by draw() and render_image(). Once the machine is offline the waiting screen (rat + "waiting for machine...") is shown again, instead of a frozen Dashboard.

Tests

• test_press_while_dark_only_wakes_backlight — first press wakes, second switches.
• test_telemetry_resume_starts_new_session — buffers reset to a single fresh sample and no phantom events on resume.
• Updated test_handle_button_press_short for the new backlight gating.

All 37 tests pass; cargo clippy clean on the simulator build.

Verification in simulator (cargo sim)

1. Start → waiting screen (rat + "waiting for machine..."). Press Down → Dashboard, terminal cleared.
2. Left/Right switch screens; Up starts a pump shot, graph fills.
3. Wait >30s with no input → UI falls back to the offline waiting screen.
4. Press Down again → new session: clean terminal, graphs restart from empty.
5. After >10s idle, the first Left/Right only wakes (no screen switch).

Note: the SPI/EMI corruption itself can't be reproduced in the simulator, but the session-start self-healing (the core fix) is exercised there.