Updated dashboard with a bunch of improvements.

[finbarrtimbers]

1. Fixed Average Batch Generation Time Bug - Removed duplicate reporting and added validation to prevent extreme
   outlier values from corrupting averages
2. Static Tokens/sec Display - Tokens per second now remain static until new batches are processed, providing more
   stable metrics
3. Improved Number Formatting - Enhanced formatting for tokens/s with proper comma separators (3,650) and K/M
   suffixes (36.5k, 3.65M) when numbers exceed thresholds
4. Training Progress & ETA - Added current training step tracking with estimated time remaining calculations based
   on training speed
5. MFU + MBU Calculations - Implemented Model FLOPs Utilization and Memory Bandwidth Utilization tracking with
   proper averaging
6. Memory Usage Metrics - Added comprehensive memory tracking including:
   - Total GPU memory usage across all actors
   - Average KV cache size per actor
   - Peak memory usage tracking
   - Integration with vLLM engine memory statistics
7. Dashboard UI Enhancements - Added new organized sections for:
   - 🎯 Training Progress (step counter + ETA)
   - 🧠 Model Utilization (MFU/MBU metrics)
   - 💾 Memory Usage (GPU memory + KV cache)
   - Enhanced styling and visual organization
8. vLLM Integration - Extended LLMRayActor with comprehensive metrics collection from vLLM engines including GPU
   memory stats and compute utilization estimates