[MSWEBX]

calloop PR: reuse the poll events buffer across dispatches

Repo: https://github.com/Smithay/calloop (fork it, branch reuse-poll-events)
Base: master at 1e9def9a3f8ce1417e559cc57a549a9516ea5aac
Patch: calloop.patch (next to this file; git apply calloop.patch on that base)
Checked on that base: cargo fmt --check clean, cargo test: 48 + 3 + 1 passed. The new test
fails on the unpatched crate (0.14.5: 1000 allocations in 1000 dispatches).

Our vendored copy (vendor/calloop, 0.14.5) carries the same change.

---

Title

Reuse the poll events buffer across dispatches

Body

Poll::poll collects each wake's ready events into a new Vec<PollEvent>, so every
EventLoop::dispatch that finds a ready source allocates (48 bytes for one event on x86-64) and
frees it again. A GUI client wakes for its display connection about once a frame, which puts one
allocation into every frame. winit's Wayland backend dispatches through calloop, so every winit
app on Wayland pays this cost. We found it while enforcing a zero-allocation steady-state frame in
a game. heaptrack showed it as the only remaining per-frame allocation besides one in
wayland-backend (Smithay/wayland-rs#TODO).

This change:

• EventLoop keeps a poll_events: Vec<PollEvent> next to its existing synthetic_events buffer.
• Poll::poll (crate-private) takes &mut Vec<PollEvent>, clears it and refills it, instead of
  returning a new Vec.
• dispatch_events iterates and drains that buffer, so its capacity survives to the next dispatch.

The level-triggered re-registration, the timer events and the error paths are unchanged. No
public API changes.

A new integration test, tests/dispatch_allocations.rs, counts allocations through a thread-local
counting global allocator. It warms up once, then dispatches a ready Generic source 1000 times
and asserts zero allocations. Without the change it reports 1000.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NwaQZQwdPnpxWEikdEfVvo