[performance] Add opt-in SDL display backend

[sayhiben]

Summary

• Add an opt-in SDL output backend without changing the default framebuffer path.
• Add DISPLAY_OUTPUT_BACKEND=framebuffer|sdl|auto and DISPLAY_SDL_DRIVER.
• Choose the SDL video driver before pygame initializes.
• Fall back to the framebuffer path if the requested SDL display cannot open.
• Document/install the Mesa EGL/GBM runtime needed for KMSDRM on the Waveshare KMS setup.

Independence

This PR now targets main directly and contains only the SDL/KMS backend and runtime-requirement docs. Benchmark-script changes remain isolated in #54.

What Changed After Bring-Up

The original KMSDRM failure was not an app rendering issue. The device had /dev/dri/card0 and SDL's KMSDRM driver, but did not have the Mesa EGL/GLES runtime installed. SDL failed with:

[Terminal] SDL display backend unavailable, falling back to framebuffer: EGL not initialized

Installing these packages fixed KMSDRM display creation on the test device:

sudo apt install libegl1 libegl-mesa0 libgles2 libgl1-mesa-dri

The Waveshare provisioning script now installs those runtime packages, and the app fallback message now points at them when KMSDRM/EGL setup is incomplete.

Measurement / Experiment Result

Measured on the Pi using #54's benchmark script in a temporary checkout, with an opt-in run:

DISPLAY_OUTPUT_BACKEND=sdl DISPLAY_SDL_DRIVER=kmsdrm   python3 scripts/benchmark_render.py --samples 5 ...

The app now opens SDL KMSDRM successfully:

/dev/dri/card0
/dev/dri/renderD128
card0-DPI-1 connected, 720x720
[Terminal] Using SDL display driver: KMSDRM

Before SDL/KMS, after the optimized framebuffer PR:

terminal_render_display_no_stream: n=3 avg=40.1ms fps=24.9
terminal_render_display_stream_active_client: n=3 avg=43.6ms fps=23.0
typing_render_chars: n=1 chars=114 avg=4842.3ms cps=23.5

With SDL/KMS:

terminal_render_display_no_stream: n=5 avg=24.5ms fps=40.8
terminal_render_display_stream_no_clients: n=5 avg=24.8ms fps=40.3
terminal_render_display_stream_active_client: n=5 avg=28.3ms fps=35.4
typing_render_chars: n=1 chars=114 avg=3772.4ms cps=30.2

Conclusion: this is a real win on the KMS-backed Waveshare device once the Mesa EGL runtime is installed. Keep framebuffer as the default for broad compatibility, but KMSDRM is now a working opt-in path and a good candidate for targeted KMS display profiles.

Validation

• python3 -m py_compile config.py display/terminal.py
• bash -n scripts/setup-waveshare-4dpi-720.sh
• Local opt-in DISPLAY_OUTPUT_BACKEND=sdl DISPLAY_SDL_DRIVER=dummy smoke test.
• Pi SDL/KMS benchmark above; tinyprogrammer.service was restarted successfully afterward.