multi-state lstm for time series forecasting

I have 100’s of distinct users and the dataset consists work/hours spent by each user over the past 1 year. Total records are over 40k. The inputs for the LSTM RNN has the categorical feature vectors corresponding to these users.The rnn should learn behavior of each user and should be able to predict the next hour spent in the per user time-series based on the past information of the same user. How to maintain separate hidden states for each user within an LSTM RNN.

I’m refering to below blog post is similar to my problem:

https://towardsdatascience.com/multi-state-lstms-for-categorical-features-66cc974df1dc

Sample data(again referring to the blog) enter image description here

My Queries:

How multi-state lstm’s can be implemented using kera?
Is it a custom lstm rnn with stateful nature?
how do we keep separate state for each user with categorical features?
Thanks in advance!