Thread `out_sharding` natively through elementary `jax.random` samplers
* Propagate `out_sharding` and inferred parameter shardings (`_sample_sharding`)
directly through `uniform`, `randint`, `normal`, `truncated_normal`,
`bernoulli`, `cauchy`, `exponential`, `gumbel`, `laplace`, `logistic`,
`pareto`, `rayleigh`, `wald`, `triangular`, and `lognormal` via `random_bits`
instead of wrapping each sampler in `maybe_auto_axes`.
* Validate broadcast-aligned parameter shardings against `out_sharding`
(raising `ShardingTypeError` on conflict), sample directly at inferred
parameter shardings when `out_sharding=None`, and preserve eager execution
under `disable_jit()` across elementary samplers (including `rademacher`,
`maxwell`, and `geometric`).
* Register `_naryop_collective_rule` on `polynomial_p` so `normal`
(`mode='high'` and `'highest'`) lowers under `jax_partitioning`.
PiperOrigin-RevId: 1020460678