Add experimental persistent BF16 full fine-tuning

[muthu813]

This PR adds an experimental persistent-BF16 parameter path for native C++ Full-FT.

Key changes:

• Persistent BF16 parameter storage
• FP32 gradients and FP32 Adam moments
• BF16-aware LayerNorm, RMSNorm, MatMul, and gradient accumulation paths
• FP32-to-BF16 SafeTensors loading
• BF16 validation tests
• GPT-2 124M full-model validation
• README documentation of the experimental extension

Validation performed on an ARM64 Android device:

• 148/148 GPT-2 124M parameter tensors loaded as BF16
• 0 FP32 parameter tensors
• Full 124M forward pass completed successfully
• 10/10 forward passes
• 10/10 backward passes
• 10/10 FP32 Adam updates
• Parameters remained BF16 throughout the 10-step run
• Separate 20-step real-data causal Full-FT experiment completed

This is an engineering extension of MobileFineTuner, not a claim of a new optimization algorithm.