Defer: nest moe_* fields into ModelConfig.moe when field count grows

[mmshad]

ModelConfig has 11 moe_* fields today (num_experts, moe_top_k, moe_frequency, moe_router, moe_shared_experts, moe_aux_loss_weight, moe_capacity_factor, moe_sequence_aux_loss_weight, moe_gradient_scale, moe_bias_schedule, moe_packed_experts). The prefix does poor-man's namespacing. Phase 14 from moe_eng_production_plan.md adds 2 more. At ~15+ fields it starts hurting ergonomics — cross-field validation in __post_init__ gets awkward, and it's harder to see what's MoE-specific vs model-general.

Eventual shape:

@dataclass
class MoEConfig:
    num_experts: int = 0
    top_k: int = 2
    # ... rest of the current moe_* fields, unprefixed

@dataclass
class ModelConfig:
    dim: int = ...
    moe: MoEConfig = field(default_factory=MoEConfig)

TOML grows a [model.moe] subtable. is_moe becomes config.model.moe.num_experts > 0. CLI overrides become --model.moe.top_k=4 — loader.py already handles arbitrary dot depth.

Not doing this now. 11 fields isn't sprawl. And _apply_dict_to_dataclass in kempnerforge/config/loader.py rejects unknown TOML keys, so flat-to-nested is a hard cutover — either rewrite every existing MoE config in the same PR or add a compat shim that rewrites moe_* under [model] into [model.moe] with a deprecation warning. Not worth paying for a cosmetic win.

Revisit when:

• moe_* field count on ModelConfig hits ≥15, or
• Cross-field validation needs to live in __post_init__ (e.g. top_k ≤ num_experts, capacity_factor > 0 requires specific routers), or
• Someone wants moe_layers: list[MoEConfig] for multi-variant composition.

Out of scope: mtp_* fields from Phase 15 (not MoE), DistributedConfig.ep (orthogonal to model arch).