test(ffi): de-flake pump_yields_when_queue_full via threshold recalibration

[G4614]

De-flake pump_yields_when_queue_full by lowering its progress >= 10 assertion — which sat right on the boundary of the ~8–10 canary ticks the bulk drain produces — to >= 5, a threshold recalibration with no production change (yield_now doesn't actually starve the canary; a worker-monopolizing producer collapses it to ~2, still caught).

Test plan

• make test:unit:rust FILTER=pump_yields → 15/15 runs pass, non-flaky.
• Two-side verified: inject std::hint::spin_loop() in place of the yield (a worker-monopolizing producer) → canary collapses → assertion fails with the named error; restore → passes.

observed	pre (>= 10)	post (>= 5)
cooperative yield_now (current prod code)	~8–10 ticks → flakes on the 9/10 edge (~40% first-try fail)	~8–10 ticks → passes with margin
worker-monopolizing producer (the real regression)	caught (~2 < 10)	caught (~2 < 5)
canary max_gap	~21 ms (no starvation)	~21 ms (unchanged)

[G4614]

The recalibration to >= 5 cuts the flake rate but doesn't fully close it on a CPU-contended host. Running the suite on an AWS box — where nextest runs the boxlite-c crate's tests in parallel, contending for the single tokio worker the canary shares with the producer — pump_yields_when_queue_full failed both nextest tries, with the canary dipping to 4 ticks:

TRY 1 FAIL: canary advanced only 6 ticks; producer likely busy-spinning instead of yielding
TRY 2 FAIL: canary advanced only 4 ticks; producer likely busy-spinning instead of yielding

With >= 5, the 4-tick run still trips the assert (the 6-tick one would have passed). Suggest lowering to >= 3: it stays clear of the ~2 busy-spin floor this PR measured, while leaving margin below the cooperative range observed here (4-6; ~8-10 on a quiet host). The discriminator is preserved (cooperative → canary advances; busy-spin → stalls at ~2) — >= 3 just sits in the gap instead of on the noisy 4-6 boundary.

(Longer term the absolute-tick assertion is inherently timing-fragile under scheduler contention; a relative baseline — e.g. assert cooperative ticks ≫ an in-run busy-spin sample — would be load-independent. But >= 3 is the minimal change to stop the flake.)