Umbrella issue of improving Orca user experience (update from time to time)

[hkvision]

[General]

• init_orca_context default cores=2, and users may not be aware of this setting and may just call it with default parameters. And in this case if user compare the local performance between the original TF/PyTorch script with the Orca version, they will find much performance drop since Orca is only using two cores while the original script will use all cores in most cases.
  Shall we change default cores to "*"?
• The setting of default OMP related environment variables, related issues:
  Default KMP settings could slow down torch estimator #4370
  Default OMP, KMP values would impact other applications #4372
• Doc for more arguments for init_orca_context #4538
• [ORCA] Torch version for ray backend #4984
• Automatically find free port for RayOnSpark #4540
• [ORCA] cannot read several files with different schemas using spark backend #4913 This may be of relatively low priority, we can probably don't support this case?

[Customized data and train]

• For non-standard/complicated train loops, users may not be able to easily migrate to Orca, related issue:
  PyTorch Ray Estimator for customized data and train loop #3557
• Some further preprocess steps after creating PyTorch DataLoader is not supported, related issue:
  Orca do not support preprocess inputs before training in forward() #4410
• Customized metrics is not supported for evaluation, for example:
  Orca only supports basic metrics, some customized metrics do not support  #4414
• Sometimes loss is computed in the forward pass and we may set the loss_creator optional? Related issue:
  yolox does not require loss creator when running on orca #4412
• PyTorch Training Operator train_batch and forward_batch can be merged to avoid duplicate code.
• For the above issues in this section, fixes can only be applied to ray and pyspark backend, bigdl backend has even less possibility to support those cases.

[Customer1 code specific issues]

• Support multiple output and multiple loss functions (MMoE for multi-task learning)
• Support PyTorch Lightning models.
• More flexibility in PyTorch train input and outputs #4448
• Orca pytorch Estimator support validation when training #4468
• Orca datatype convert issue #4476

cc @jason-dai @shane-huang