lotka_volterra + stoch_vol MCLMC recipes are stale-invalid: warmup runs away at their own configs on current blackjax

Found by the blackjax #969 golden guard (see the fix PR on blackjax). Both committed recipes' warmup stages, re-run at their own configs (n_warmup=500, seed=42) against the CURRENT pinned blackjax, produce step-size runaway — not the committed values:

| Recipe | Committed ss | Pinned blackjax re-run | blackjax-#969-fix re-run |
|---|---|---|---|
| lotka_volterra mclmc_tuning | 0.30 | 1220 (runaway) | ≈0 (collapse — see blackjax#973) |
| stoch_vol mclmc_lrd_tuning | 4.24 | 737 (runaway) | 0.152 (recovered; LRD geometry identical) |

The recipes were generated at blackjax 1.6.dev84, before the #969 trap manifested at these seeds. They cannot be re-blessed under either version as-is:
- **stoch_vol**: re-cert under the #969 fix is plausible now (recovered warmup, identical LRD geometry) — candidate for straightforward re-cert once the fix merges.
- **lotka_volterra**: its NaNs appear position/solver-caused (ODE model), so the fixed adaptation collapses instead of converging — re-cert is gated on blackjax#973's outcome (divergence-response design), or on a model-side init/parameterization change.

The other 5 committed mclmc-family recipes verified bitwise-identical under the fix.

— 🤖 Blackjax-devs AI TL

Start by rerunning the lotka_volterra and stoch_vol MCLMC recipes with n_warmup=500 and seed=42 against the current pinned blackjax, then compare the warmup step sizes with the committed values. Re-certify stoch_vol after blackjax#969 merges; for lotka_volterra, wait for blackjax#973 or a model-side initialization or parameterization decision before considering it done.

Written by the indexing model from the issue text.