Rewrite snapshot v2: JSON-only, no cloudpickle, Rust tokenizer

Eliminates cloudpickle dependency entirely. New approach:
- Tokenizer: save_pretrained → load via Rust tokenizers lib (0.2s)
- Model config: JSON manifest (not pickled class objects)
- Tensors: mmap'd from safetensors files (0.07s)
- Model: reconstruct from config on meta device + load_state_dict

Benchmark results (Qwen2.5-7B, 15.2GB, RTX 4090):
  Warm hydrate: 0.32s (was 0.42s with cloudpickle)
  Cold hydrate: 1.40s (excluding torch import + inference)

  Cold breakdown:
    manifest:     0.02s
    Rust tok:     0.21s
    mmap tensors: 0.07s
    import model: 1.08s (transformers.models.qwen2)
    meta+load_sd: 0.01s

V1 cloudpickle snapshots still loadable via _hydrate_v1 fallback.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: lightsofapollo