v0.27.0: Graph functions; Improved Go backend (fusion ops); Quantization dtypes; more ML layers & fixes

• Package backends: major refactoring to add support for functions/closures.
  • Added backends.Function, which now holds all the "ops" methods.
  • Added NewFunction, Closure and Call.
  • Renamed backends.Op -> backends.Value.
  • Added FusedOps, allowing backends to expose fused (more efficient) operations -- with proper/automatic
    fallback to decomposed operations when not supported or for gradients.
  • Added ErrNotImplemented error and IsNotImplemented(err) function.
  • Added Quantization struct, QuantizationScheme (Linear, NF4), and NF4LookupTable.
  • Removed Dot() operation (redundant with DotGeneral).
  • DotGeneral() now takes a DotGeneralConfig struct, with options for setting the accumulator and output dtypes.
• Package simplego:
  • Added Float16 support (thx @timkaye11)
  • Added dedup of computation nodes (aka. "common subexpression elimination" CSE) (thx @timkaye11, @janpfeifer)
    • ~6% speedup for CSI-Adult demo training.
  • DotGeneral: Pre-blocking of the blocked path, which may lead to deduplication of blocking nodes (@timekaye11).
  • DotGeneral: Added smallMatMul execution path, optimized for small matrix multiplications (thx @timkaye11)
  • Experimental packgemm support leveraging simd operations (@ajroetker, @janpfeifer)
  • Funtions/closures support (thx @ajroetker)
  • Added Reverse operation.
  • Added fused operations: FusedGelu, FusedDense, FusedSoftmax, FusedLayerNorm,
    FusedScaledDotProductAttention, FusedAttentionQKVProjection.
  • Added FusedQuantizedDense: fused dequantization + matmul + bias + activation for Int4/Int8 weights
    with Linear and NF4 quantization schemes, block-wise scales, and optional zero points.
  • FusedScaledDotProductAttention: added ScaledDotProductAttentionConfig options struct with
    QuantizedMatmuls flag for optional uint8 quantized Q@K/attn@V matmuls (awaiting go-highway
    release for actual acceleration).
  • Bitcast refactored to pure bit reinterpretation; sub-byte unpacking moved to ConvertDType.
• New package bucketing:
  • Tools to manage bucketing of tensors (or anything else) -- thx @ajroetker
• Package dtypes:
  • Added 'Uint2', 'Uint4', 'Int2', 'Int4'.
• Package graph:
  • Added Function concept (and support for closures) and the Function.Call operation.
  • Control Flow: Added While and If operations.
  • Order operations: Added Sort, SortFunc, TopK, BottomK.
  • Fixed Bitcast for packed sub-byte types: Int4, Int2, Uint4 and Uint2, so they can be "bitcast"
    back and forth from/to uint8 (bytes), to ease quantization.
  • Added Atan2 function.
  • Added test helper functions to test various backends at once.
  • Fixed Gather validation of indexVectorAxis to check against startIndices rank instead of operand rank.
  • Exec graph compilation is now concurrent, avoiding redundant compilations for the same graph shape.
• Package ml/layers/attention: Improved MultiHeadAttention; Added KVCache support.
  • Added Grouped Query Attention (GQA) support.
  • Added UseQKVProjection() option for fused Q/K/V Dense projection.
  • FusedScaledDotProductAttention now supports boolean masks.
• Package ml/layers/attention/pos: Added PositionalEncoder interface, and "RoPE" (Rotary Positional Encoding) implementation.
• Package ml/models/transformers:
  • Added a Transformer "model": a collection of transformer layers are setup based on given configuration.
• Package ml/decode:
  Added a Decoder object to generate text given a sequential model.
• Package ml/decode/sample:
  • Added implementation of various sampling strategies (greedy, temperature, beam-search, top-k, top-p, etc.), used by the decode package.
• Package ml/layers/activations:
  • Added HardSwish.
• Package examples:
  • Separated in its own sub-module, to separate its dependencies.
  • Added gpt2: A simple GPT-2 implementation using the new transformers and decode packages. It downloads the model from HuggingFace.
  • Added textgen: a minimal transformer text generation model that can be trained.
  • Added gemma3: A simple Gemma 3 implementation using the onnx-gomlx package to convert the model, and go-huggingface to download the model and run the tokenizer.
  • Added mxbai-rerank: A cross-encoder reranking example using the MixedBread Reranker v1. It uses the onnx-gomlx package to convert the model, and go-huggingface to download the model and run the tokenizer.
  • Added BERT-base-NER: A BERT-base model fine-tuned for Named Entity Recognition.
• Bumped github actions versions to the new "Node24" ones.