fix: avoid possible data leak while rendering

[alekseysotnikov]

Between separate read and the reset another thread can add scripts/dirty-ids/full-render flags that get silently overwritten ("leaked").

swap-vals! atomically reads the old value and sets the new, closing the race window.

[rschmukler]

@alekseysotnikov thanks for catching this - it was a real bug.

I ended up going a step further and replacing the atoms + semaphore with a single LinkedBlockingQueue (hyper.render.queue). The swap-vals! fix is correct per-atom, but with three separate atoms there's still a cross-atom inconsistency window (e.g. reading full-render-needed* as false, then another thread sets it to true before we read pending-partials* - we'd do a partial when a full was requested). It's self-healing since the semaphore wakes the loop again, but the queue approach just makes the whole class of problem go away. This refactor makes it so we effectively don't have any shared mutable state between threads.