ci: install jax on the recovery job so the gradient-path fits run

[wshlavacek]

Follow-up to #518 / #517.

The recovery job added in #518 was dispatched for the first time to size it (the point of shipping it dispatch-only). It came back failure at ~11.7 min: 125 passed, 2 failed — and the two failures were a config gap in the job, not a stochastic tolerance flake:

FAILED tests/test_real_world_examples.py::test_real_world_runs_through_bngsim[Rijal-2025/lacud5_ode]
FAILED tests/test_real_world_examples.py::test_real_world_runs_through_bngsim[Rijal-2025/five_dl1_ode]
  PybnfError: Gradient assembly needs jax to transform a log-scaled parameter's
  gradient into sampling space ... the optional 'jax' extra.

Those two real-world cases fit log-scaled parameters through jax-backed gradient assembly, and they raise rather than skip when jax is absent. The recovery job was built with setup-pybnf's default jax: 'false', so it failed deterministically — it will fail identically every run until jax is installed.

Fix

Add jax: 'true' to the recovery job — exactly as tests.yml's pytest-jax job already does for the same gradient path. A complete recovery environment provisions jax.

Verification

• The two previously-failing cases pass locally in ~15s with jax present.
• Full end-to-end confirmation: this branch was dispatched via workflow_dispatch on the branch ref to confirm the whole recovery tier goes green under jax before merge. (Result linked in comments.)

Sizing takeaways from the first dispatch (context for #517)

• Recovery tier ≈ 12 min on a 2-core runner — the ">1 hr" concern was a dev-machine artifact; it's comfortably schedulable.
• Zero stochastic tolerance flakes this run — the only failures were the deterministic jax gap fixed here.
• Side observation: the pre-existing slow job ran 32 min, vs the ~8 min its comment claims. Not touched here; flagged for a possible follow-up.