Chronos: Embrace bigdl-nano in Chronos

[TheaperDeng]

Background

Currently Chronos rely on orca.automl for their pytorch model single node support. This will cause

• Future chronos-inference will rely on orca (and dllib) which make the whole package fat.
• Can not use nano's speed up.
• pytorch-lightning Trainer will bring more feature to Chronos (such as efficient lr/bs tuning)

Implementation Step

We will implement the whole plan mainly in 5 PRs

• add a new nano.pytorch.pl_base_model, this model should add time series specific training step ...

• add chronos.metric to replace orca.automl.metric. Chronos: add independent chronos.metric #3608

• change forecaster(distributed=False) path to bigdl-nano(pl) Chronos: pytorch based forecasters on bigdl-nano #3539

• change TSPipeline path to bigdl-nano(pl) Chronos: TSPipeline on bigdl-nano #3609

• change Automodel (except .fit) to bigdl-nano(pl) Chronos: AutoProphet & AutoARIMA without orca when inference #3745

[jason-dai]

1. distributed=false: using nano (possibly with pseudo-distributed mode)
2. distributed=true: using orca

[TheaperDeng]

1. distributed=false: using nano (possibly with pseudo-distributed mode)
2. distributed=true: using orca

yes, that is the plan