docs: add Halo fine-tuning guide for LFM MoE

[advpropsys]

Summary

Adds lfm/fine-tuning/halo.mdx, a guide to training LFM MoE models with Halo, White Circle's open-source distributed training framework. Halo adds multi-GPU and multi-node parallelism to the native Hugging Face workflow while models remain standard transformers modules and checkpoints remain loadable with from_pretrained.

The guide:

• explains how one LFM integration and YAML structure carry across SFT, LoRA, GRPO, and environmental GRPO
• documents FSDP, EP, TP, ETP, and EP+TP support for LFM MoE models
• includes the published single-B300 LFM2.5 comparison: 31,262 vs 28,330 tok/s/GPU at 8K and 35,517 vs 29,520 at 16K, against Axolotl
• provides a runnable EP2 SFT path for LFM2.5-8B-A1B
• links to the complete Halo LFM2 cookbook for larger models, LoRA, inference, and RL
• adds Halo to the fine-tuning overview and navigation

Validation

• npm run snapshot:check
• verified the external Halo, White Circle, and model links