perf(acrylic): release blur render targets while the effect is skipped

Turning blur off (blur 0), switching to low-spec mode or enabling
reduced transparency after the acrylic targets existed left the capture
FBO and both blur ping-pong targets allocated for the rest of the
session, although nothing draws with them while the effect is skipped.

They are now released at a frame boundary and rebuilt on demand:
- ImGuiAcrylic::ReleaseIdleBuffers() runs right after app.preFrame(),
  before NewFrame(), when the previous frame's draw data has been
  rendered, and once in the resize-watcher frame (GL and DX11). While
  the effect is skipped it frees the targets; when the effect is wanted
  again it reallocates them once, with no per-frame retry if that fails;
- resize() while skipped only records that a reallocation is needed;
- the release checks the GL names (and each DX11 texture) directly, so a
  half-allocated ping-pong pair is freed too; DX11 also resets the
  ping-pong index;
- the capture paths and applyBlur() return early without a capture
  buffer, which would otherwise blit the default framebuffer onto
  itself;
- each release logs "released targets (effect skipped)".

With blur on, the path and the output are unchanged (screenshot sweeps
within noise). After re-enabling, the first frame draws the tinted
fallback until a fresh capture exists.

Measured with software GL, toggling every 30 frames for 1200 frames: the
targets were released exactly one frame after each switch-off (22 of 22)
and rebuilt on each switch-on (21 of 21), with no GL or framebuffer
errors. While off at 1280x720, with a fixed malloc mmap threshold, the
process holds 4.62 MB less (the 1280x720 RGBA8 capture plus two 320x180
RGBA16F blur targets) and RssAnon is about 3.1-3.5 MB lower. With
glibc's default threshold the freed memory stays in the heap and RssAnon
does not visibly drop.

Known issue: every off/on pair now frees and reallocates those targets.
Under software GL (llvmpipe) with the default glibc allocator the freed
texture memory is not reused: toggling every 2 frames grew RssAnon from
49 to 906 MB over 240 pairs, every 8 frames from 56 to 179 MB, and every
30 frames it levels off near 120 MB. It is not a GL object leak (with a
fixed mmap threshold the mapped bytes stay flat), and on a hardware
driver (WSLg d3d12) RssAnon stayed between 67.5 and 74.9 MB under the
2-frame churn (94.3 MB without the release). Users reach fast toggling
by holding Ctrl+Shift+Down (key repeat, about 20 Hz) or by dragging the
blur slider back and forth across 0. A short delay before releasing
would avoid it if software-rendered users matter.

Not verifiable here: the DX11 release and reallocation at runtime, and
macOS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/main.cpp

@@ -1538,6 +1538,8 @@ int main(int argc, char* argv[])
         SDL_GL_MakeCurrent(ctx->window, ctx->gl_context);
         ImGui_ImplOpenGL3_NewFrame();
 #endif
+        // Free/rebuild acrylic targets at the frame boundary (after MakeCurrent on GL: FBOs are per-context).
+        dragonx::ui::effects::ImGuiAcrylic::ReleaseIdleBuffers();
         ImGui_ImplSDL3_NewFrame();
         ImGui::NewFrame();
 
@@ -1951,6 +1953,8 @@ int main(int argc, char* argv[])
         // Pre-frame: font atlas rebuilds and schema hot-reload must
         // happen BEFORE NewFrame() because NewFrame() caches font ptrs.
         app.preFrame();
+        // Acrylic targets are freed/rebuilt here, between frames: nothing recorded can reference them.
+        dragonx::ui::effects::ImGuiAcrylic::ReleaseIdleBuffers();
 
         // Smooth font-scale: compensate visual font size via FontScaleMain
         // when the atlas hasn't been rebuilt yet (during slider drag).