fix: tear down the component context tree on close, freeing it by refcount

[maartenbreddels]

Summary

Closed render trees could not be freed by reference counting: use_state setters, reacton.ipyvue.use_event handlers, and _event_handler_exception_wrapper all capture the _RenderContext and ComponentContext strongly in closures that live in the hook state and on widget CallbackDispatchers (where event handlers deliberately stay after close, to avoid syncing _events right before the comm closes). That makes the whole closed tree one reference cycle — rc → hook state / widgets → closure → rc — which only a generation-2 gc can free.

On a solara server this shows up as a memory sawtooth whose peaks scale with per-kernel state: closed kernels (including e.g. use_memo payloads) pile up until the collector's own schedule kicks in. Measured with a task-heavy test app: idle footprint bounced up to 231 MB; with the tree refcount-clean it converges to a ~170 MB plateau (peak 174 MB). Details, the measurement harness, and the gc-disabled diagnostic that found these exact chains: widgetti/solara#1177 (docs/memory-usage-inspection.md) and widgetti/solara#1176.

The fix

Capture both contexts via weakref.ref in the three closure factories. Behavior during the tree's life is unchanged (the render context keeps the tree alive, so the weakrefs always resolve). After the tree is collected the closures become inert:

• a setter no-ops — there is no tree left to update;
• the exception wrappers re-raise instead of appending to a dead context — previously the error was swallowed into a tree nobody could see.

Testing

• New test_no_reference_cycles_after_close (4 parametrizations, incl. a use_state + on_click component): asserts a closed tree is freed with gc disabled. Red on master (3/4 fail) → green with the fix. Passes with REACTON_FAST=1 too (make_setter and the wrappers are shared with the fast render context).
• Full suite: 171 passed; the 5 failures + 1 error in generate_test.py / test_set_state_with_dataframe are pre-existing on master in the same environment (pandas/codegen version drift).
• End-to-end (solara + patched reacton): a browser-driven kernel exercising reactive/computed/task/use_task + a 4.5 MB use_memo payload is fully freed by refcounting after the page closes, with gc disabled.

🤖 Generated with Claude Code

[maartenbreddels]

Design changed after checking history (thanks Maarten for the pointer): the first push used weakrefs in the closures, which is exactly what cf61378 reverted in Nov 2023 ("a setter might be the only reference" — weak captures sometimes lost live state in solara). The current version keeps strong captures and instead empties the component context tree at close(), which achieves the same refcount-collectability without that failure mode. Same red→green test, same end-to-end result.

[maartenbreddels]

The readthedocs failure was unrelated config drift: readthedocs dropped the ubuntu-20.04 build image, so every new build fails config validation (the last green docs builds, #49/#51, predate the deprecation). Added a one-line commit bumping to ubuntu-22.04.