Reference issue

Closes gh-21833
(I'm not aware of anything else that must be transitioned, and I think gh-21921 can be considered separate.)

What does this implement/fix?

The _transition_to_rng decorator was helpful, but not all our cases matched the usual pattern, so I thought it would be helpful to add a standard test for all the SPEC 7 transition work.

During the process, I found two small issues:

• The scipy.sparse.random_array documentation has always stated that it uses the RandomState singleton if if random_state is not provided, but it actually uses an unseeded Generator (for speed). I have opted to leave the documentation alone and change the implementation; this adds a small incentive to transition to the new rng argument. @dschult, if you don't want to go this route, please add a commit to this branch: undo the change to random_array, pass replace_doc=False into its _transition_to_rng decorator factory, manually adjust the documentation, and change the tests as necessary.
• When multivariate_normal_qmc and multinomial_qmc received RandomState via the random_state keyword, we were getting an error when this was passed along to Sobol via the rng keyword. The simple fix is to pass them to the Sobol via the random_state keyword, instead. Note that I'm only using the random_state keyword if the user has already done so, so I think that's OK; if we get duplicate DeprecationWarnings when those get turned on, we can deal with it then. (We're branching soon.)

I also found a small issue with the adjustments to PermutationMethod and BootstrapMethod before adding these tests. The gist of it is that we need to keep track of whether the user passed in the argument as random_state or rng so that the call to permutation_test or bootstrap does the right thing with it. For simplicity, we also probably shouldn't allow rng to be mutated during the transition. It's just an extra thing to worry about, so I removed the setter.