malloc_trim(0) was called at the end of every update() and view(),
reaching 60-200 Hz during typical scrolling, resize, or animation.
Each call walks the glibc heap (10 us to several ms depending on
fragmentation) and could consume a substantial fraction of the frame
budget in worst cases.
Throttle trim() to once per second using a thread-local Instant,
preserving the existing API. RSS stays bounded (1 Hz is enough to
release collectable pages soon after) while per-frame cost becomes
a single thread-local check plus a duration comparison.
No call-site changes required; the three existing trim(0) invocations
in src/app/cosmic.rs (update, view multi-window, view single-window)
now fall under the throttle transparently.
To reduce compile-times and avoid some overhead to binary size, this will modify some of our
generic functions to use non-generic inner functions where possible. The inner functions are
marked carefully with `#[inline(never)]` to prevent being inlined by LLVM at their callsites
While looking for generic functions to optimize, I have also taken the opportunity to annotate
public non-generic getters and setters with `#[inline]` to ensure that LLVM will inline them
across crate boundaries. By default, only generic functions are automatically inlined, and
only when enabling fat LTO are constant functions reliably inlined across crate boundaries.
