reduce restrictions for setting optimizer/dataset of a model

[kparichay]

The public API requires the optimizer/dataset to be set before ml_train_model_compile. However, these restrictions can be relaxed and both can be later right until the training is being done:

1. dataset - current implementation allows this already, only documentation changes are needed
2. optimizer - minor implementation changes are needed (remove the check for the model is compiled when setting the optimizer and move the initialization of the optimizer to the training itself)

Thank you @pkosko for pointing these out.

[taos-ci]

:octocat: cibot: Thank you for posting issue #1849. The person in charge will reply soon.