[Dev] Add Llama3 training example and fix cache save & static shape marking

[wtr0504]

🗂️ PR Category

• ✨ New Feature
• 🚀 Optimization (performance, memory, etc.)
• 💥 Breaking Change
• 🐛 Bug Fix
• 🛠️ Development / Refactoring
• 📚 Documentation
• 🧹 Chore (Dependencies, CI/CD, Configuration, etc.)
• 🧪 Testing

📝 Description

Summary

• Add end-to-end Llama3 training example (example/training/) with FSDP support, a distributed training script, and an Nsys profiling launch script.
• Fix a cache save bug where aot_autograd artifacts were empty, causing compiled graphs to fail to persist correctly.

Changes

• example/training/llama3.py — Llama3 model definition adapted to use magi_compile
• example/training/train.py — distributed training loop with FSDP and NVTX profiling hooks
• example/training/train.sh — torchrun launcher with optional Nsys profiling
• magi_compiler/magi_backend/piecewise_compiler.py — workaround for empty aot_autograd artifacts on cache save
• magi_compiler/utils/nvtx.py — profiler for iteration