Chronos: support validation_data and validation loop in forecasters' fit

[TheaperDeng]

time series data might not be that large and overfitting has become a problem for many users. We should bring a validation loop to forecaster's fit and make sure the best ckpt with lowest validation loss is used after the fit.

• [x] behavior definition
• [ ] local forecaster implementation 1 (tcn, s2s, lstm, nbeats)
• [ ] local forecaster implementation 2 (autoformer, mtnet, tf2 forecasters)
• [ ] distributed forecaster implementation (tcn, s2s, lstm, nbeats)