perf(app): spread pinned workers across physical cores before SMT siblings

[gg582]

Problem

cwist_app_pin_worker() pins worker i to the (i mod n)-th allowed CPU in ascending order. That only spreads workers across physical cores when the kernel numbers SMT siblings far apart (0,N). Many hosts number them next to each other (0,1, 2,3, …), including many cloud VMs and some Intel and ARM boxes. On those hosts, CWIST_WORKERS=4 puts the four workers on the two hyperthreads of two cores while the remaining cores sit idle.

Follow-up to the #319 investigation. The benchmark box there numbers siblings 0,6, so it never showed this.

Change

src/sys/app/worker_affinity.h:

• Each allowed CPU gets a rank: the number of its allowed SMT siblings with a lower ID. Siblings come from topology/core_cpus_list, falling back to thread_siblings_list on older kernels.
• CPUs are picked in (rank, ID) order, so every physical core gets one worker before any core gets a second.
• Unchanged cases: no SMT, 0,N numbering, no sysfs (some containers), or an unreadable or malformed topology file. In all of these the order stays the old ascending one, so the change is a strict improvement or a no-op.
• The inherited mask is still respected. A core with only one allowed thread counts as rank 0.

Siblings	Before (workers 0..7)	After
0-1,2-3,4-5,6-7	0 1 2 3 4 5 6 7	0 2 4 6 1 3 5 7
0,4 1,5 2,6 3,7	0 1 2 3 4 5 6 7	0 1 2 3 4 5 6 7
4-way 0-3,4-7	0 1 2 3 4 5 6 7	0 4 1 5 2 6 3 7

There's no new API or env knob, and no public header changes.

Tests

tests/test_worker_affinity.c points the reader at a fake sysfs root and checks:

• adjacent, split and 4-way orders (the last one through thread_siblings_list only);
• a partial mask;
• malformed files, which fall back to ascending order.

The existing mock and real-child cases are unchanged; they run with no topology exposed.

On this host (12 CPUs, siblings 0,6) the real order is unchanged: 0 1 … 11. taskset -c 0,1,6,7 gives 0 1 6 7.

Local runs of the test are clean, including ASan/UBSan, and clang-format reports nothing.