[Tracking Issue] Relax training M0 migration and polishment

[SiriusNEO]

In this fork @Ubospica and me collaborated on the development of a training workflow on relax. We finished M0 and the efforts included:

• An automatic differentiation pass on relax high level operators.
• Some relax operators for training.
• Gradients registered for some relax operators.
• A framework supports training with different optimizers.
• A trainer wrapper.

Now these results are being migrated to new struct info infra. During this migration, we also try our best to polish our previous work. Related PRs and Progress:

• (Dependency) Operator Legalizer
  • Legalizer V0 [Transform] Operator legalizer V0 #96
  • Legalize more operators [Transform] Legalize some operators #101
• (Dependency) TOPI collapse_sum patch [TOPI] Expose the interface of topi.collapse_sum #102
• Some related operators
  • collapse_sum_to, collapse_sum_like [Op][Manip] collapse_sum_like, collapse_sum_to #87
  • cross_entropy, log_softmax, nll_loss [Op][NN] cross_entropy, log_softmax, nll_loss #94
  • exp [OP] Unary tensor operator R.exp #100
  • abs (For L1Loss) [Op] Completing the unary operators and refactor test #113
• Operator Gradients
  • Gradients in new infra. Gradients for log_softmax and nll_loss, Gradients for split and concat. collapse_sum erasure. [Op] Migration: Gradients for some operators #98
  • Gradients for nll_loss and conv2d [WIP][Relax][Op] Register gradient for nll_loss and Conv2d #114
• Automatic Differentiation Pass
  • CopyWithNewParams utils [Relax] Add CopyWithNewParams interface #91
  • AD Pass (Some detailed checkpoints can be found in this PR) [Relax][Training] Add automatic differentiation pass #103
• Optimizer & Trainer APIs
  • Optimizer [Relax][Training] Optimizer API #107
  • Refactor on Optimizer and Gradient [Relax][Training] Refactor Optimizer and Gradient #121
  • Loss functions [Relax][Training] Loss functions #112
  • Trainer [Relax][Training] Trainer API #115