GL native-wide (16:9) compositor: port SW wide path to OpenGL backend

Mirrors gpu_sw_renderer.c's native-wide model on the GL backend. Canonical
VRAM (the hr FBO) stays 100% faithful; native-wide lives in SEPARATE per-base_x
wide FBOs, each framebuffer-targeting primitive mirrored at local x = vram_x -
base_x + OFFSET. Present reads the displayed buffer's wide FBO via the CPU
present path (Option B: lowest-risk reuse). 4:3 / non-opted games never call
wide_configure, so the wide path never runs and the canonical pass is identical.

- GEO_VS/TEX_VS: add u_xoff (x translation) + u_xhalf (x clip half-extent);
  x term is now (a_pos.x+u_shift+u_xoff)/u_xhalf-1.0. Defaults 0/512 set once at
  init, restored after every wide pass, so canonical reduces to (x+shift)/512-1.
- Wide FBO management (WIDE_MAX_SURF=4): lazy RGBA8 color tex + FBO per base_x;
  glb_wide_configure/set_target/disable_target/clear mirror the SW functions.
- Double-draw: each draw func emits its canonical pass first (unchanged), then if
  g_wide_cur re-issues the SAME glDrawArrays into the wide FBO via shared
  wide_target_begin/end helpers (geometry / shaded+textured: both STP passes with
  identical blend + u_semipass). Flat-rect full-screen-overlay extend replicated
  (covers [0, native_w) -> fills the full wide width) via a dedicated full-width
  wide pass with the generic per-tri mirror suppressed.
- glb_render_wide_display: glFinish then glReadPixels(GL_BGRA, GL_UNSIGNED_BYTE)
  yields ARGB8888 byte-identical to sw_render_wide_display / upload_present_tex;
  rows flipped (GL bottom-left origin) and alpha forced 0xFF to match SW.
- Vtable: wire the 5 .wide_* ptrs so gr_wide_supported() returns nonzero for GL.
- main.cpp present: guard the VRAM-FBO early-return with !wide_present so a wide
  GL frame falls through to the existing CPU readout block; all other paths same.

Builds clean. Visual validation pending (human).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

runtime/src/main.cpp

@@ -1335,7 +1335,12 @@ static void sdl_vblank_present(void) {
          * the PSX_GL_FORCE_CPU_PRESENT diagnostic sync the FBO down and use
          * the CPU readout below. */
 #ifndef PSX_SDL_NO_RENDER
-        if (g_gl_active && g_gl_fbo_present && !di.depth24) {
+        /* Native-wide GL present: the wide compositor lives in separate per-base
+         * FBOs that the deterministic VRAM-FBO present can't read, so a wide
+         * frame falls through to the CPU readout path below (gr_render_wide_
+         * display -> sdl_pixel_buf -> gl_renderer_present). 4:3 / FMV / non-wide
+         * frames keep the fast VRAM-FBO present (early return) unchanged. */
+        if (g_gl_active && g_gl_fbo_present && !di.depth24 && !wide_present) {
             gl_renderer_present_vram((int)di.display_x, (int)di.display_y,
                                      (int)present_w, (int)h, g_video_aa ? 1 : 0,
                                      fmv_frame ? 1 : 0);