Add TornadoVM transformer and GPU execution optimizations

Introduced extensive performance improvements for TornadoVM, including numerical, memory, and GPU execution optimizations. Enhancements include quantized weight support, vectorized operations, key-value caching, parallelized attention, and optimized GPU memory transfers. These changes significantly improve efficiency for transformer-based models like LLaMA.

Author: mikepapadim

README.md

@@ -444,6 +444,68 @@ The secret sauce that transforms regular Java code into GPU-accelerated compute
 
 -----------
 
+## TornadoVM Transformer Optimizations
+
+### Core Numerical Optimizations
+- **Quantized Weight Support**
+  - Optimized implementations for Q8_0 and Q4_0 formats
+  - Block-based quantization with FP16 scale per 32-element block
+- **Vectorized Matrix Operations**
+  - Uses vector parallelism with configurable unroll factors
+  - Processes 4 elements at once with vectorization
+- **Loop Unrolling**
+  - Strategic unrolling for performance (16x factor in matrix operations)
+  - Reduces branch penalties and improves instruction-level parallelism
+- **Fused Multiply-Add (FMA)**
+  - Uses fused operations for better numerical precision and performance
+  - Optimizes dot product calculations
+
+### Memory and Caching Optimizations
+- **Key-Value Cache**
+  - Efficiently stores past key-values for autoregressive generation
+  - Organized by layer, position, and dimension for fast access
+- **Scale Caching**
+  - Avoids redundant decompression of quantized weights
+  - Caches scale factors for efficient block processing
+- **Optimized GPU Memory Transfers**
+  - Minimizes host-device data movement
+  - One-time transfer of static data (weights, caches)
+  - Per-execution transfer of dynamic data (position, activations)
+- **Device-to-Device Data Consumption**
+  - Efficient data transfer between operations
+  - Reduces PCI-E bandwidth bottlenecks
+
+### Algorithmic Optimizations
+- **Parallel Reduction RMS Normalization**
+  - Implements two-phase reduction for efficient normalization
+  - Work group optimization for parallel sums
+- **Rotary Position Embeddings (RoPE)**
+  - Optimized implementation for positional encoding
+  - Efficient rotation of query and key vectors
+- **Optimized Float16 Decoding**
+  - Fast decoder for half-precision floating point format
+  - Special case handling for better performance
+- **Parallelized Attention**
+  - Computes attention heads in parallel
+  - Optimized softmax with max subtraction for numerical stability
+- **Fused Feed-Forward Networks**
+  - Combines operations for SwiGLU variant used in LLaMA models
+  - Optimized SiLU and GELU activation functions
+
+### GPU Execution Optimizations
+- **Layered Execution Planning**
+  - Organizes computation as separate layer-based task graphs
+  - Strategic scheduling of operations
+- **Work Group Optimization**
+  - Tailored worker grid configurations for different operations
+  - Matches GPU hardware characteristics
+- **Local Memory Optimization**
+  - Strategic use of local/shared memory for reductions
+  - Optimizes bandwidth-intensive operations
+
+-----------
+
+
 ## Early performance of v1.0
 
 ![GPULlama3.java Performance Comparison](./docs/performance.png)