Fix numba CauchyRV loc/scale

[notluquis]

Closes #2308.

The numba dispatch computed (loc + z) / scale instead of loc + scale * z, so the sampled variable had location loc/scale and scale 1/scale.

-        return (loc + rng.standard_cauchy()) / scale
+        return loc + scale * rng.standard_cauchy()

rng_fn_scipy and the JAX dispatch were already correct, so only numba was affected — but it is the default linker. logp is untouched, so NUTS posteriors were fine; anything that draws was not, including sample_prior_predictive on a HalfCauchy prior (HalfCauchyRV.rv_op builds abs(cauchy(0, beta))).

Tests

test_unaligned_RandomVariable is the only test checking these samplers against their distribution, and it has three cases. Two used a degenerate scale:

case	was	now
cauchy	scale=1.0	scale=3.0	catches the bug above: p=1.7e-08 → passes
gumbel	scale=1.0	scale=4.0	no bug, but was blind
t	scale=np.pi	unchanged	already fine

At scale=1, (loc ± z) / 1 is algebraically identical to loc ± 1 * z, so both the location and scale errors vanish together and Cramér–von Mises passes.

GumbelRV has a hand-written core too (loc - scale * np.log(-np.log(U))) and is correct, but was equally undetectable. Verified by injecting the analogous slip: the new case fails at p=7.2e-08 with it, passes without.

Audit

Checked every hand-written numba core (t, halfnormal, pareto, invgamma, gumbel, wald) plus the default-path RVs against scipy with non-degenerate parameters. All match — cauchy was the only real bug.

tests/link/numba/test_random.py: 84 passed, 2 xfailed.

Note for reviewers reproducing this: PyTensor reuses the cached compiled graph across source edits, so a naive before/after run shows both variants passing. The numbers above use a separate base_compiledir per run.