v0.4.0: API change (onnx package refactored); Many new ops and models support; Fused ops;

[janpfeifer]

This release include the new v0.4.0 and v0.3.[56] (never actually released) versions.

v0.4.0: More ops and fused ops; Added support for model Florence-2; onnx package refactored.

• Package onnx: split implementation into internal/onnxgomlx/..., onnx is now just a public API.
  • API change: it's a small change but onnx.Model is now an interface (it was a pointer to an object)
    and the constructors reside in onnx/parser.
• Mod operator: Supports both fmod=1 (C-style, sign follows dividend) and fmod=0 (Python-style, sign follows divisor) with broadcasting and dtype promotion
• onnxImplicitFloatPromotion for float-only ops (Sqrt, Exp, etc.).
• Concat dtype alignment: When dtype promotion is enabled, all Concat operands are cast to the first operand's dtype, preserving Int64 for shape/index tensors.
• isVariableConstant loosening: Float variables with "const" in the name are now accepted as materializable constants (needed when Concat dtype promotion casts Float32 constants to Int64).
• Sub-graph name shadowing fix: convertSubGraph now saves and restores parent entries in nodeOutputToNode / variableNameToValue instead of unconditionally deleting them on cleanup.
• convertIf() rework: Uses GoMLX's native If with closures instead of the Where-based approach.
• Added support for Florence2 model;
• Updated GoMLX dependency to v0.27.0.

v0.3.5: New ops for various models support (Gemma, Snowflake,CLAP, etc); Many fixes and improvements.

• Added quantized fusion patterns for dense layers, QKV projections, and scaled dot-product attention (SDPA). (by @ajroetker)
• Fixed isZeroInitializer to handle tensors with zero-sized dimensions (e.g., [batchSize, 0]).
• Add ONNX operators for Gemma and Snowflake Arctic models: ReduceL2, SimplifiedLayerNormalization (RMSNorm),
  RotaryEmbedding, MultiHeadAttention.
• Simplified optional DType auto-promotion.
• Added Resize operation for "CLAP" model support.
• New Einsum op support (2 operands).
• Various bug fixes across multiple ops and sub-graph handling.
• Added ReduceMax, ReduceMin, ReduceSum, ReduceProd, and NonZero ONNX operations.
• Optional auto-promotion of dtypes in case of mismatches: this is an error
  in ONNX specification, but some Pytorch models do this. See Model.AllowDTypePromotion().
• Added support for reading variables from external data files (due to proto 2/4 Gb size-limit)
• Several fixes (see logs)

---

This discussion was created from the release v0.4.0: API change (`onnx` package refactored); Many new ops and models support; Fused ops;.