fix(wg_mgr): yield 1 tick per peer in disco_periodic_probes

[fudio101]

What

disco_periodic_probes() iterates every active peer with no yield point in
between. On a large tailnet the full loop can run 50-170ms back-to-back;
since wg_mgr shares core 1 with coord (see CLAUDE.md's task table),
this starves coord (and anything else pinned to core 1) long enough to
trip watchdogs on real hardware. Adds vTaskDelay(1) at the top of each
per-peer iteration — gives other same-core tasks scheduler time without
measurably affecting DISCO cadence (the loop still completes within the same
outer 1s cycle, just with a shorter peak burst length).

Source

Adapted from cplewes/microlink
9f6af750 (not cherry-picked — base has diverged too far). cplewes's
original commit message cites a specific incident (BLE-RPC notify pool
starvation on a long-uptime capture); this fork doesn't have that
downstream's BLE stack, so the comment here is adapted to reference our own
core/task assignment instead.

Related issues

Closes #24

Independent of #47/#48 — touches ml_wg_mgr.c only, branched fresh off
main.

Test plan

idf.py build isn't runnable in this sandbox (no ESP-IDF installed).
Manual check: confirmed vTaskDelay(1) sits inside the per-peer loop body,
at the top of each iteration (matches cplewes's placement — delays before
the first peer too, which is intentional and harmless). Needs hardware
re-test on a tailnet with enough peers to actually observe the starvation
this addresses; a 1-2 peer test tailnet won't exercise it.