Make soak test actually detect leaks

[ausimian]

Summary

The soak test as it landed in #5 had two problems that only showed up under the full suite:

1. In-suite drift swamped the signal. Running standalone it passed cleanly, but running after the other 101 tests MLX's Metal buffer pool kept ~2 tensors' worth of arena alive past `clear_cache`. At 4 MB tensors / 2 MB tolerance the test failed on legitimate allocator behaviour.
2. Single-tensor detection was never achievable. Arena drift scales with tensor size (~2 tensors, regardless of per-size), so any single-tensor retention is indistinguishable from the noise floor.

Re-tuned around rate detection instead of absolute retention:

• 1 MB tensors × 2000 iterations (still ~1 s wall time)
• 4 MB tolerance — absorbs the ~2 MB of Metal arena drift, leaves headroom for one or two stray retentions
• A per-iteration leak rate scales linearly with `@iters`; arena drift does not. With 2000 iters, a 1 % leak rate accumulates to ~20 MB — 5× the tolerance.

Validated by injection: a deliberate 1 %-per-iter leak yields delta 45 MB vs 4 MB tolerance → fails by 10×. Clean path passes in ~1 s as part of `mix precommit`.

Test plan

• `mix precommit` green locally (102 tests including soak).
• Injection test (retain every 100th tensor) fails cleanly with informative diagnostic.
• CI precommit green on this PR.