Change some mechanism

[kaylode]

• Use global iteration index now, instead of epoch index
• Generalize base trainer
• Change dataloader mechanism to work with custom collator function, batch sampler. Usage example:

dataloader:
    train:
      name: DataLoaderWithCollator
      args:
        batch_size: 32
        drop_last: true
        shuffle: false
        collate_fn:
            name: MixupCutmixCollator
            args:
              mixup_alpha: 0.4
              cutmix_alpha: 1.0
              weight: [0.2, 0.2]
        sampler:
            name: BalanceSampler
            args: