fix(flet-charts): prevent unbounded browser memory growth in MatplotlibChart

[FeodorFitsner]

Summary

• Replace the generic flet.canvas.Canvas + capture() flow used by MatplotlibChart with a dedicated MatplotlibChartCanvas widget that composites matplotlib diff frames in CPU memory. Holds at most one ui.Image at a time, avoiding the Picture.toImage allocations that accumulate on Flutter web (CanvasKit/WASM) and aren't reclaimed by the JS GC.
• Without this fix, the matplotlib animation example grew to ~22 GB browser memory in 2–3 minutes and crashed the tab.

What's in the PR

New widget

• MatplotlibChartCanvas (Python control + Flutter widget) with apply_full(bytes) / apply_diff(bytes) / clear() invoke methods. Decodes PNGs, keeps a Uint8List backbuffer, composites diff frames where alpha != 0, and shows the result via a single ui.Image rebuilt with decodeImageFromPixels.

Backend / chart wiring

• MatplotlibChart switches from fc.Canvas shapes + capture() to the new widget. Rubberband (zoom selection) becomes a positioned Container overlay in a Stack.
• TimerFletAsyncio subclass captures the main asyncio loop at construction so FuncAnimation keeps working when _timer_start() is invoked from a worker thread (used by our threaded render path).
• Render runs on asyncio.to_thread on native runtimes so the asyncio loop stays free for pointer events. Pyodide / WASM falls back to inline rendering with aggressive yields, since to_thread runs synchronously there.
• threading.Lock serializes worker-thread renders against main-thread figure.savefig to prevent the print_figure manager=None race that otherwise crashes Download.
• Receive loop coalesces stale binary frames and stale draw JSON requests so pan/zoom doesn't visibly play back buffered motion.

flet core (flet.canvas.Canvas) image disposal

• Dispose old _capturedImage / shape _image / intermediate Picture and Codec after replacement (post-frame for safety).
• Gapless playback in drawImage: keep the cached image visible while new bytes decode, instead of going blank for a frame.

Safari fix

• Defensive Uint8List.fromList(bytes) copy before instantiateImageCodec. Safari's WASM runtime can free buffers across async boundaries and otherwise throws EncodingError: Loading error.

Test plan

• examples/extensions/charts/matplotlib_chart/animate in Edge — memory stays flat, animation smooth.
• Same in Safari — no EncodingError, memory flat.
• examples/extensions/charts/matplotlib_chart/toolbar — pan/zoom smooth, Download saves the file.
• Native desktop run — animation, interaction, Download all work; Download blocks briefly on render-in-flight (expected).

Summary by Sourcery

Replace MatplotlibChart’s generic Canvas-based rendering with a dedicated image-stream canvas to stop unbounded browser memory growth and keep interactions responsive across platforms.

New Features:

• Introduce MatplotlibChartCanvas widget to display WebAgg-style matplotlib image streams with full and diff frame support.
• Expose MatplotlibChartCanvas and its resize event through the flet-charts Python API and Flutter extension registry.

Bug Fixes:

• Prevent unbounded browser memory growth from Picture.toImage allocations during matplotlib animations, especially on Flutter web.
• Fix race between worker-thread renders and main-thread figure.savefig that could crash downloads.
• Ensure Safari no longer throws image decoding errors when rendering matplotlib charts.
• Avoid visible playback of buffered pan/zoom motions by coalescing stale render frames and draw requests.
• Eliminate blank-frame flicker when rapidly updating Canvas images by keeping the previous image visible until the new one is ready.
• Dispose previously captured images, intermediate pictures, codecs, and cached Canvas images to avoid leaks.

Enhancements:

• Run matplotlib renders on worker threads on native runtimes while keeping the main asyncio loop free for input handling, with a safe fallback path for Pyodide/WASM.
• Add a TimerFletAsyncio implementation so matplotlib FuncAnimation timers can be started from worker threads without event loop errors.
• Move Matplotlib rubberband selection from Canvas drawing to a Flutter overlay for simpler composition.
• Serialize matplotlib drawing operations with a lock to coordinate threaded rendering and main-thread operations like saving figures.

[cloudflare-workers-and-pages]

Deploying flet-examples with    Cloudflare Pages

Latest commit:	204e353
Status:	✅  Deploy successful!
Preview URL:	https://efbdb9cd.flet-examples.pages.dev
Branch Preview URL:	https://web-memory-crash.flet-examples.pages.dev

View logs

[sourcery-ai]

We've reviewed this pull request using the Sourcery rules engine

[cloudflare-workers-and-pages]

Deploying flet-website-v2 with    Cloudflare Pages

Latest commit:	204e353
Status:	✅  Deploy successful!
Preview URL:	https://fd08f26a.flet-website-v2.pages.dev
Branch Preview URL:	https://web-memory-crash.flet-website-v2.pages.dev

View logs