feat: implement MultiDtypeGenerator with FX Graph passes

- Add MultiDtypeGenerator for generating low-precision (float16/bfloat16/float8) samples
- Implement FX Graph passes for dtype conversion (no AST/string manipulation)
- Preserve BatchNorm/LayerNorm buffers as float32 for numerical stability
- Add autocast metadata for runtime mixed-precision support
- Thread-safe extraction with proper locking
- Comprehensive error handling and device detection
- Validated on CV (ResNet18) and NLP (BERT) samples

Architecture:
- graph_net/torch/multi_dtype_generator.py: Main generator with RunModelDecorator
- graph_net/torch/multi_dtype_passes/: FX Graph pass implementations
  - pass_base.py: Base class for dtype conversion passes
  - dtype_conversion_pass.py: Concrete dtype conversion implementation
  - autocast_wrapper_pass.py: Autocast metadata injection
  - pass_mgr.py: Pass manager

Features:
- Real FX Graph passes (not string replacement)
- Smart weight preservation (BN/LN buffers stay float32)
- Automatic device type detection
- Compatible with test_compiler and Agent workflows
- Extensible filter system for unsupported operations

Fixes:
- Correct val_map semantics in FX Graph rewriting
- Proper dtype detection (avoid false positives on int tensors)
- CPU fallback when CUDA unavailable
- Import organization and error handling

Author: Dayuxiaoshui