Implement a `split_rngs` mechanism, like in `flax.linen.scan`

I have the following problem: I want to optimize/find the fixed point of a function that uses flax's [`Dropout`](https://flax.readthedocs.io/en/latest/guides/dropout.html) and therefore under the hood it uses [`make_rng`](https://flax.readthedocs.io/en/latest/notebooks/flax_sharp_bits.html).
What this means, is that when the function is called multiple times inside the optimizer/fixed point solver, the dropout mask will change between calls.

For recurrent neural networks whose cells use dropout, this is fixed in flax using a `split_rngs` mechanism available in the [`scan` function](https://flax.readthedocs.io/en/latest/api_reference/_autosummary/flax.linen.scan.html).
Namely, the doc says:
> `split_rngs` – Split PRNG sequences will be different for each loop iterations. If split is False the PRNGs will be the same across iterations.

The same is available for the [while loop](https://flax.readthedocs.io/en/latest/api_reference/_autosummary/flax.linen.while_loop.html).

In my case, I would basically like to do something like `split_rngs={"dropout": False}`.

I think the best way to tackle this would be to implement special while loops for flax cases in [here](https://github.com/google/jaxopt/blob/main/jaxopt/_src/loop.py). Happy to discuss the API and whether this feature makes sense.

My basic use case is to implement Deep Equilibrium Models where the fixed point defining function uses dropout.