Use proper Poisson sampling in documented examples and keras_api

[ryan112358]

The JAX library core library includes APIs and Implementations of batch selection strategies, including standard Poisson sampling. However, some of our example binaries do not use this, and instead load in the data sequentially as fixed size batches, which is common practice without privacy. This is a problematic flaw with the implementations that invalidates the privacy guarantee and should be fixed.

An example that does correct sampling is given in dp_logistic_regression.py.

Task:

• Update keras_api.py so that Poisson sampling is used instead of deterministic fixed-batch sampling. This may require imposing additional constraints on the input dataset format. In particular, the dataset needs to support efficient random access (in-memory arrays, or arrayrecord format).

[RamSaw]

Neerajpathak07 could you comment on this issue, so that I could assign it to you?

Debangan Ghosh, Neeraj created the first PR, it is already under review, we might need to close your PR. We've updated our contribution guidelines to avoid race conditions like that, please take a look. I will try to find time to look at your PR and maybe if there are nice ideas, merge them from your PR (I might need to ask you to sync and adopt the code a bit).