a refactored modeling_act for cpu and memory optimization

[eDeveloperOZ]

Performance Optimization: Replace .to() Operations with Buffer Management

What this does

This PR optimizes tensor operations in the ACT Policy implementation by replacing unnecessary .to() device transfers with proper buffer management. The changes focus on improving memory efficiency and reducing CPU overhead while maintaining training performance.

Label: ⚡️ Performance

Key changes:

1. Added registered buffers for frequently used tensors
2. Pre-computed tensors during initialization
3. Implemented buffer templates for repeated operations
4. Optimized memory allocations using expansion instead of creation

How it was tested

The changes were thoroughly tested through:

1. Full training runs comparing baseline vs optimized implementation
2. Performance metrics monitoring showing:

• ~50% reduction in CPU utilization (from ~2% to ~1%)
• More stable memory usage pattern around 2000MB
• Maintained training convergence and loss curves
• Consistent or improved data loading times

All standard tests pass without modification as functionality remains unchanged.

Performance Results

Key metrics show significant improvements:

# CPU Usage
# Before: ~2% average utilization
# After:  ~1% average utilization

# Memory Pattern
# Before: Variable usage with frequent fluctuations
# After:  Stable ~2000MB with smoother allocation pattern

# Training Performance
# - Loss curves maintain expected convergence
# - Data loading times: No degradation
# - Update times: Consistent with baseline