deps: unpin jax/jaxlib now that numpyro 0.21.0 ships fix

[conorheins]

Summary

• Reverts the temporary <0.10 cap on jax/jaxlib (and jax[cuda12]) introduced in deps: pin jax/jaxlib <0.10 pending numpyro fix (#389) #391 across pyproject.toml and setup.cfg.
• Removes the inline TODO comments referencing pyro-ppl/numpyro#2173.
• numpyro#2173 merged 2026-05-02 and shipped in numpyro 0.21.0 the same day. The fix wraps the xla_pmap_p import in a try/except so the absence on jax ≥ 0.10.0 no longer breaks import numpyro (and therefore pybefit).

Verification

• Confirmed numpyro 0.21.0 release tag contains the try/except guard in numpyro/ops/provenance.py.
• uv lock --dry-run resolves cleanly: jax 0.10.0, jaxlib 0.10.0, numpyro 0.21.0, pybefit 0.1.23.
• Standard Test CI passes on this PR.
• Manual Branch Nightly run on this branch passes.

Notes

This closes out the workaround from #389. Should be a clean inverse of #391.

🤖 Generated with Claude Code

[conorheins]

First nightly run (25428706834) green on the steps that gate this PR — Install dependencies and Run nightly-marked tests both pass on jax 0.10.0 + numpyro 0.21.0, confirming the unpin's primary goal.

It surfaced one secondary regression in examples/advanced/pymdp_with_neural_encoder.ipynb: 5 jnp.clip(..., a_min=…) calls that fell off jax 0.10's deprecation cliff (TypeError: clip() got an unexpected keyword argument 'a_min'). All 6 reported notebook "failures" were that single root cause cascading through later cells (NameError on variables that were never assigned because the prior cell crashed). Library code is unaffected — grep for a_min=/a_max= across pymdp/ is clean, only the one notebook used the deprecated kwarg.

Fixed in 225c9a1 by switching to numpy-2.0-style min=. Nightly passes again with this fix