improve performance of find_stack_level

[MarcoGorelli]

This has been improved in pandas, so bringing the improvement over https://github.com/pandas-dev/pandas/pull/45247/files

On master:

In [1]: from numpyro.distributions.discrete import PRNGIdentity

In [2]: %%timeit
   ...: PRNGIdentity()
   ...: 
   ...: 
<magic-timeit>:1: FutureWarning: PRNGIdentity distribution is deprecated. To get a random PRNG key, you can use `numpyro.prng_key()` instead.
3.64 ms ± 70.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Here:

In [1]: from numpyro.distributions.discrete import PRNGIdentity

In [2]: %%timeit
   ...: PRNGIdentity()
   ...: 
   ...: 
<magic-timeit>:1: FutureWarning: PRNGIdentity distribution is deprecated. To get a random PRNG key, you can use `numpyro.prng_key()` instead.
5.68 µs ± 58.8 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

[fehiepsi]

Awesome!!! Thank you for the improvement!