[Module] Modularize MoE components

[acisseJZhong]

Stack from ghstack (oldest at bottom):

• -> [Module] Modularize MoE components  #2571

Motivation

Continues the modularization effort from #2428 (Embedding), #2434 (RMSNorm), and
#2527 (Linear). This PR makes MoE components (GroupedExperts,
TokenChoiceTopKRouter, TokenReorderer) inherit from Module and use the
Config/build() pattern, enabling per-component configurability and consistent
weight initialization.

Summary

• GroupedExperts inherits from Module with a nested Config dataclass. dim,
  hidden_dim, and num_experts use field(init=False) so they are set at build() time
  by the parent MoE.
• TokenChoiceTopKRouter inherits from Module with a nested Config dataclass.
  Replaces the flat gate_bias: bool field with gate: Linear.Config, enabling full
  gate configuration (e.g., Linear.Config(bias=True)). dim and num_experts use
  field(init=False).
• TokenReorderer inherits from Module with a no-op init_weights (stateless
  module). num_experts and top_k use field(init=False).
• MoE.Config replaces flat router/expert fields (top_k, score_func, route_norm,
  route_scale, gate_bias, use_grouped_mm, num_expert_groups, num_limited_groups,
  _debug_force_load_balance) with nested experts: GroupedExperts.Config and router:
  TokenChoiceTopKRouter.Config.
• Updates all model configs (DeepSeek V3, Qwen3, GPT-OSS, Llama4) to use nested
  router/expert configs.
• Adds 34 unit tests covering Config/build(), init_weights, forward passes, and
  nested config propagation for all modularized components.

cc @fegin #2540 may needs to rebase on this.

[tianyu-l]

Maybe better if it doesn't come with a default, so that model config constructor have to specify.

[tianyu-l]

makes sense. I wonder if we should put the following to router as well

• tokens_per_expert and expert_bias buffer to router module
• load_balance_coeff to router config