New XGBoost strategy: cyclic training

[yan-gao-GY]

Issue

Description

The exiting bagging strategy for XGBoost cannot perform client-by-client (cyclic) training, although cyclic training is a popular FL fashion.

Proposal

This PR achieves cyclic training with two steps:

• Override the sample() method of client_manager() to fix the order of selected clients
• Create a new strategy fedxgb_cyclic, and override configure_fit() and configure_evaluate() to do sequential client selection.

This way, there is only one single client selected to participate in the training per round with the fashion of client1-server-client2-server-clientN.
The trained local XGBoost trees will be passed to the next client as an initialised model for next round's boosting.

[charlesbvll]

It might be good to add some details about the difference between bagging and cyclic, and/or why a user would prefer to use one over the other. Otherwise it looks good!

[yan-gao-GY]

@charlesbvll thanks for your suggestion and error-catch! I have offered an explanation of bagging and cyclic in readme in the later PR. The docstring has been updated.