[kacy]

lock_sweep/spawn became bimodal at #1226: some runs took about 160 ms and others about 280 ms. The instruction count rose by only 0.3%. With one worker the time was unchanged. At 16 workers, user CPU went from 2.0 s to 3.4 s. That pattern is cache-line contention, not extra work.

#1226 added statics, which moved the shared stack-pool mutex onto the same cache line as SCHEDULER. Every scheduler call reads SCHEDULER, so each lock of the pool evicted that line from every core. Inside the scheduler, searching (a compare-and-swap on every spawn) and the injector queue sat next to the workers and slab pointers. Adjacent Worker entries also shared lines, so a thief locking one worker's queue evicted its neighbour's.

This change puts the stack pool, the scheduler, searching, the injector and each worker on their own 64-byte lines.

Interleaved medians at 16 workers on cores 0-7,16-23, main against this branch:

row	main	this branch
lock_sweep/spawn	255 ms	155 ms
green_fanout	1085 ms	885 ms
lock_sweep/breaker	146 ms	142 ms

The other concurrency rows are unchanged. chan_fanout lands anywhere from about 70 to 130 ms in both builds.