Refactor scene transitions to reduce memory usage and improve cleanup

[bubbadubc]

This refactors Stencyl's built-in scene transitions to reduce unnecessary full-screen bitmap allocations, remove redundant scene redraws, and make transition-owned rendering resources cleanly disposable.

The main issue was that several transitions created one or more full-screen BitmapData objects and repeatedly redrew the scene into them during the transition. At high resolutions this becomes expensive very quickly, and it also interacts poorly with shader-based rendering because the final shader-composited image already exists on the GPU.

The fix:

• Added TransitionSnapshot.hx as a shared screenshot/capture path.

• On native OpenGL/OpenGLES targets, transition snapshots are copied directly from the already-rendered backbuffer into a GPU texture using copyTexSubImage2D rather than redrawing the scene into a CPU BitmapData.

• If GPU capture is unavailable, TransitionSnapshot falls back to the existing BitmapData.draw() behavior so non-GPU targets remain supported.

• TransitionSnapshot explicitly disposes both its BitmapData wrapper and any GPU texture it creates.

• Crossfade now uses the shared snapshot implementation instead of owning its own full-screen capture logic.

• Slide no longer maintains multiple full-screen scene buffers or recomposites them every frame. It keeps only the scene state required for the transition.

• Circle/Spotlight and Rectangle no longer require a full-screen screenshot. Their reveal areas are generated directly as transition geometry over the live scene.

• Bubbles no longer captures the full screen into BitmapData. The transition operates on the live scene using expanding circular regions.

• Pixelize/Blur removes the two full-resolution BitmapData buffers and the per-pixel CPU processing loop. It now renders into one reusable low-resolution buffer and scales that result with nearest-neighbor filtering.

• Blinds, Fade In, and Fade Out remain lightweight vector transitions, with cleanup tightened so owned graphics and tween references are released consistently.

• Public transition constructors, directions, and timing behavior remain unchanged.

At 3840x2160, a single 32-bit full-screen BitmapData is roughly 31.6 MiB. Several of the previous transitions allocated multiple buffers of this size, so this removes a significant amount of avoidable transition memory usage.

The goal of this change is not to introduce a new rendering system. It keeps the existing transition API and behavior while using the already-rendered GPU frame when available, falling back safely where necessary, and making resource ownership and cleanup explicit.

TODO / Known limitation:

• Bubbles and Spotlight still use more memory during the transition than I would ideally like.
• Their current peak usage is not worse than the existing implementation; it is roughly in line with the previous behavior.
• The current refactor removes the unnecessary full-screen bitmap capture/redraw paths, but these two transitions still have a relatively expensive rendering cost because of their animated circular geometry/masking.
• This is not considered a blocker for the refactor, but there is still room to investigate whether their temporary rendering cost can be reduced further.