Enable CUDA provider option configuration for C#

[hariharans29]

Description:

This is the CUDA EP equivalent of the PR #7808 (All the design choices have already been discussed there).

This change makes the OrtCUDAProviderOptions struct opaque and supports setting options that can be converted to a string. Just like the TensorRT PR, an API to set options that can't be strings (user_compute_stream and default_memory_arena_cfg) is not yet supported and will be added incrementally. Since these are not very commonly used, it is okay to leave them unsupported for now. For users who still need them, they can still use the legacy struct and its corresponding ORT APIs.

Motivation and Context

1. Enables basic configuration of the CUDA EP via C# (Resolves C# adding option for cudnn_conv_algo_search : DEFAULT #9730 and How to apply "gpu_mem_limit" to CUDA Execution Provider in C#? #8995)

2. Introduces a new C API CUDA EP configuration - cudnn_conv_use_max_workspace. This was already available via Python but not available for C API users. This configuration is particularly important for fp16 Conv heavy models. See Resnet - converted Onnx model is 2.9X slower than pyTorch model in V100 gpu #9162 and [Doc] Performance tuning of convolution heavy models with the CUDA EP #9916 for details as to why. By this reasoning, this PR should resolve this perf complaint - Why Float16 model takes twice as long as Float under GPU #9420.

3. Addresses a backlog item to make the OrtCUDAProviderOptions opaque

[pranavsharma]

LGTM 👍