Correctness fixes from ROOT's `tmva/sofie` that are missing in this fork

[guitargeek]

Summary

This fork shares its operator classes, shape inference and RModel code generation with SOFIE in ROOT, and its GPU code is generated by the same classes. Correctness bugs fixed upstream therefore also affect the GPU path here, unless they are ported.

I think before developing further features in the SOFIE GPU version, it is urgent that all bugfixes from ROOT are also applied in the GPU path, so users don't get frustrated by re-discovering the same bugs when trying to run their ONNX models on the GPU with SOFIE.

All missing fixes as a list of PRs

Some PRs contain several of the commits above.

Confirmed missing:

• [tmva][sofie] Fix Pool ceil_mode ignored in ShapeInference and Generate root-project/root#21776: Pool ceil_mode
• [tmva][sofie] Fix Conv/Pool padding & ceil_mode bugs, and harden the SOFIE test suite root-project/root#22862: AveragePool divisor
• [tmva][sofie] fix SAME_UPPER/SAME_LOWER autopad padding formula root-project/root#21793: SAME_UPPER/SAME_LOWER autopad (Pool part)
• [tmva][sofie] Extend ONNX operator and type support, fix two codegen bugs, speed up Reduce and TopK root-project/root#23331: Reduce index arithmetic and TopK with sorted=0 (TopK part)
• [tmva][sofie] Fix the handling of tensors read by several nodes root-project/root#23471: tensors read by several nodes
• [tmva][sofie] Fix initialized tensors used as model outputs root-project/root#23483: initialized tensors used as model outputs
• [tmva][sofie] Fix Range output-size expression for symbolic INT64 inputs root-project/root#23200: Range with symbolic INT64 inputs
• [tmva][sofie] Fix Conv+Add fusion null path root-project/root#21979: Conv+Add fusion null path
• [tmva][sofie] Implement padding generation for ConvTranspose with explicitly set output_shape root-project/root#23088: ConvTranspose padding
• [tmva][sofie] Fix the shapes and the types of Gemm, Reshape, Clip, Tanh and Range root-project/root#23486: shapes and types of Gemm, Reshape, Clip, Tanh and Range

Probably missing:

• [TMVA][SOFIE] Fix Memory Bloat in ROperator_Pool by removing unused _xpad allocation root-project/root#23082: Pool memory bloat
• [tmva][sofie] Add new operators (ScatterND, Clip, GELU, ...) and several fixes root-project/root#21621: Expand, Gemm, Shape, Slice, Where and MLPF parsing fixes (also in needs review)
• [tmva][sofie] Fix NonZero to define max output shape value in the Session constructor root-project/root#21458: NonZero maximum output shape
• [tmva][sofie] Fix big endian root-project/root#22723: big endian

Needs review:

• [tmva][sofie] Add support for optimal memory allocation of dynamic tensors  root-project/root#20434: GNN execution order and dynamic tensor fixes
• [tmva][sofie] Implement support for dynamic tensor shapes  root-project/root#18633: dynamic shape handling
• [tmva][sofie] fix compute strides for output tensor in topK and add input tensors for memory optimization in GEMM root-project/root#19023: TopK strides, Gemm memory optimization
• [tmva][sofie] Correctly add intermediate tensors for ROperator GRU root-project/root#18645: GRU intermediate tensors
• [tmva][sofie] Add optimization modes in SOFIE root-project/root#18355: bool tensor allocation
• [tmva][sofie] New Keras Parser from SOFIE root-project/root#20871: PyMVA Keras 3 and LayerNormalization

[guitargeek]

Closed by #59 . Thanks @sanjibansg !