v0.2.1: Pipeline Training Mode & GAN-Only Training

🚀 Major Features

Pipeline Training Mode (NEW)

Multi-step sequential training with per-step configuration:

• Sequential stages: Define warmup → GAN → flow in single YAML config
• Per-step freeze/unfreeze: Selective component training
• Loss-threshold transitions: Automatically progress when loss reaches target
• Inline optimizer/scheduler configs: Different optimizers per step
• Per-step checkpoints/metrics: Step-specific files for focused analysis
• Full resume support: Mid-pipeline resume with state preservation

GAN-Only Training Mode (NEW)

• train_reconstruction parameter (default: true)
• Set to false for adversarial-only training
• Use case: SPADE conditioning without reconstruction overhead

Enhanced Logging & Monitoring

• Batch timing (Xs/batch)
• Step-specific progress files
• Correct GAN loss keys (loss_gen, loss_disc)
• Mid-epoch sample generation with batch numbers

🐛 Bug Fixes (12 commits)

• Fixed encoder gradients for GAN-only mode
• Fixed VAE trainer calling for GAN-only
• Fixed EMA creation for GAN-only mode
• Fixed checkpoint resume state tracking
• Fixed diagram generation for pipeline metrics
• Fixed sample file renaming

📚 Documentation

• docs/PIPELINE_ARCHITECTURE.md (547 lines)
• Pipeline training in docs/TRAINING_GUIDE.md (100+ lines)
• Pipeline testing in CONTRIBUTING.md
• Complete CHANGELOG.md entry

🧪 Testing

• 446/446 tests passing ✅
• All linters clean ✅

📦 Installation

pip install fluxflow-training==0.2.1

🔗 Dependencies

Requires: fluxflow-core v0.2.1+

Full Changelog: v0.1.1...v0.2.1