AMLDisplay: force the modeset when frame packing starts or stops

[cinema-ONE]

Frame packing is the same mode name with a doubled timing. aml_set_drmDevice_active() only performs
the modeset when force_mode_switch is set or the mode name differs from the current one, so when the
display is already at 1080p24hz - the previous title was 1080p24, or a 3D title in a playlist follows
another - a frame-packed title sends 3dfp to the driver without a modeset. The driver keeps the 2D
timing: vinfo stays at 1080 lines, no VSIF is sent, and a frame-packing projector shows the top eye
stretched with no 3D.

Force the modeset when the stereo mode switches into or out of HARDWAREBASED. m_stereo_mode is
UNDEFINED before the first switch, so the kernel's stereo_mode is read for the comparison, as
handle_display_stereo_mode() already does.

Measured on an Ugoos AM9 Pro (S905X5M, hdmitx21) against a JVC DLA-RS4100 through an HDFury VRROOM:

• nightly 20260904: playlist TAB clip (1080p24) -> MVC clip: display/mode 1080p24hz, stereo_mode 8,
  vinfo height 1080, 3d_info 0; projector "1080p 24", top half of the frame stretched, no 3D. Same
  clip started from 2160p25: height 2205, 3d_info 2, projector "1080p 24(FP)", correct 3D.
• with this change (same playlist): height 2205, 3d_info 2, and back to height 1080, 3d_info 0
  on stop.

videoplayer.adjustrefreshrate = "On start/stop" hides the problem, because every stop returns to
the desktop mode and the next start is a real modeset; "On start" and playlists expose it.

🤖 Generated with Claude Code

[Portisch]

The forced mode switch is set here: https://github.com/CoreELEC/xbmc/blob/aml-5.15.196-22.0/xbmc/windowing/amlogic/WinSystemAmlogicGLESContext.cpp#L180

[cinema-ONE]

Thanks - that block is the right place to look, and it explains the gap. It only reaches the stereo-mode
comparison (line 179) when the resolution is considered unchanged, and that test compares
iWidth/iHeight. For frame packing the RESOLUTION_INFO carries the doubled iHeight (2205), so a
switch from a 1080p24 TAB/SBS/2D mode into frame packing is classified as a resolution change, the
branch is skipped, m_force_mode_switch stays false, and aml_set_drmDevice_active() then sees the
same mode name and does nothing. The driver gets 3dfp without a modeset and keeps the 1080-line timing.

Reproduced today on the stock nightly 20260905 (9b78f0024d) on an Ugoos AM9 Pro, Adjust display refresh rate = On start, video playlist HTAB clip -> MVC clip, debug log:

21:14:21.999 CWinSystemAmlogicGLESContext::CreateNewWindow: m_bWindowCreated: true, frac rate 1(1), hdrType: none(none), force mode switch: no
21:14:21.999 CWinSystemAmlogicGLESContext::CreateNewWindow: cur: iWidth: 1920, iHeight: 1080, iScreenWidth: 1920, iScreenHeight: 1080, fRefreshRate: 23.98, dwFlags: 04, nativeGUI: true
21:14:21.999 CWinSystemAmlogicGLESContext::CreateNewWindow: res: iWidth: 1920, iHeight: 2205, iScreenWidth: 1920, iScreenHeight: 1080, fRefreshRate: 23.98, dwFlags: 34, nativeGUI: true
21:14:22.201 CAMLDisplay::handle_display_stereo_mode setting new mode: 3dfp
21:14:22.202 CAMLDisplay::set_display_resolution: try to set mode: 1080p24hz (FramePacking)
21:14:22.202 CAMLDRMUtils::aml_set_drmDevice_active - mode 1080p24hz is already set

/sys/class/display/mode stays 1080p24hz, stereo_mode 8, vinfo height 1080 (yesterday's
measurement, same sequence); the projector shows the top half of a flat picture. A stop between the two
clips hides it: the stereo-off transition re-runs the whitelist search and the desktop mode comes back,
so the next start is a real modeset.

If you would rather fix it in CreateNewWindow, comparing iScreenWidth/iScreenHeight in the
same-resolution test (and/or forcing when (current_resolution.dwFlags ^ res.dwFlags) & D3DPRESENTFLAG_MODEMASK is non-zero) would reach the existing stereo check. I kept the change in
AMLDisplay::set_native_resolution() because that is where the mode name and the 3D mode meet, but I am
happy to rework the PR whichever way you prefer.

Analysis and text prepared with Claude Code (Anthropic); the measurements were run on the reporter's box.