Feature/207 kllama gguf - #307

Implements: #207

# Conflicts:
#	settings.gradle.kts

Implements: #207

Implements: #207

Implements: #207

…pping

- Create GGUFTokenizer that extracts vocabulary from GGUF metadata
- Implement BPE encoding with greedy merge algorithm
- Handle byte fallback tokens (<0xXX> format) and SentencePiece markers
- Make tokenizer path optional for GGUF models in CLI
- Update JVM and Native CLI to use embedded tokenizer when available

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Configure shadowJar task to create kllama-all.jar with all dependencies
bundled. Run with: ./gradlew :skainet-apps:skainet-kllama:shadowJar

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Remove Format enum and format parameter from LlamaWeightLoader
- Remove loadFromKarpathyBin() and all Karpathy-specific helper methods
- Update loadLlamaRuntimeWeights functions to remove format parameter
- Simplify CLI to only support .gguf files with embedded tokenizer
- Fix Q8_1 dequantization bug (bytesPerBlock was 40, should be 36)
- Update tests to use GGUF format tensor shapes

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Outlines the path to making KLlama the #1 choice for multiplatform
LLM inference, competing with jlama while offering unique KMP capabilities.

Key phases:
- Performance: mmap loading, quantized kernels, SIMD, KV-cache
- Developer Experience: Chat API, streaming Flow, templates, tools
- Model Support: Mistral, Phi, Gemma, Qwen, MoE architectures
- Platform Acceleration: Metal, WebGPU, NNAPI
- Distribution: Maven Central, CocoaPods, npm, documentation

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Documents the core architecture for enabling:
- Zero-copy mmap loading across all platforms
- Pluggable compute backends (CPU, MLX, Metal, WebGPU, CUDA)
- Unified memory support on Apple Silicon
- Block-wise quantized inference without full dequantization

Key abstractions:
- TensorData: Heap, Mapped, Quantized implementations
- TensorStorage: Host vs Device memory management
- ComputeBackend: Platform-specific operation execution
- ImportHint: Zero-copy vs lazy vs device-copy strategies

This architecture enables running 70B models in 16GB RAM with GPU
acceleration and zero memory copies on Apple Silicon.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Implement memory-mapped file support for loading GGUF models without
copying tensor data into heap memory. This enables efficient loading
of large models by leveraging OS-level lazy loading and page sharing.

New components:
- FloatBufferTensorData interface for buffer-backed tensor access
- MmapFloatTensorData implementation wrapping MappedByteBuffer
- MmapTensorSource factory for creating tensor views from mmap regions
- MmapLlamaLoader for zero-copy GGUF loading (F32 tensors)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Update competitive comparison table: Memory-mapped I/O now ✅ (JVM)
- Mark Phase 1.1 Memory-Mapped GGUF Loading as COMPLETED
- Update Key Gaps section to reflect mmap progress

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

BitNet uses ternary weights {-1, 0, +1} enabling matmul with
additions only (no FP multiply). This provides 5-10x speedup
potential and is a unique differentiator.

Phase 1 updated priorities:
- 1.1 Memory-mapped GGUF ✅ COMPLETED
- 1.2 BitNet/Ternary support 🆕 HIGH PRIORITY
- 1.3 Standard quantized kernels (Q4_K, Q8_0)
- 1.4 JVM Vector API integration
- 1.5 KV-Cache optimization

Existing BitNet foundation:
- Ternary DType already defined
- DenseTernaryTensorArray for 2-bit packed storage
- GGUF TQ1_0/TQ2_0 enum values present

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Add dequantTQ1_0 and dequantTQ2_0 functions in LlamaWeightLoader
- Create TernaryTensorData interface and Ternary2BitTensorData implementation
- Implement addition-only TernaryMatmul kernel (no FP multiplies)
- Add comprehensive unit tests for dequantization and ternary ops
- Update quant_format.md with TQ documentation
- Mark BitNet support as implemented in roadmap

The ternary matmul enables efficient inference for BitNet-style models
where weights are constrained to {-1, 0, +1}, replacing multiplications
with conditional additions/subtractions.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>