Unconditionally re-enable /W4 on Windows and fix warnings which showed up

[mtavenrath]

Description

I've set the compiler warning level for msvc back to /W4 for the CUDA backend to ensure better pre-commit testing and fixed the warnings in the core onnxruntime & CUDA EP which showed up after this change.

Motivation and Context

The warning level on gcc/clang is higher than the one on MSVC when enabling the CUDA backend resulting in CI potential failures when filing a PR developed on Windows. Increase the warning level to have more overlap in warnings between MSVC and gcc/clang.

fix issue #19565

[tianleiwu]

/azp run Linux GPU CI Pipeline, Linux GPU TensorRT CI Pipeline, Windows GPU CI Pipeline, Windows GPU TensorRT CI Pipeline, orttraining-amd-gpu-ci-pipeline, orttraining-linux-gpu-ci-pipeline, Linux MIGraphX CI Pipeline, Big Models

[azure-pipelines]

Azure Pipelines successfully started running 8 pipeline(s).

[snnn]

/azp run Linux GPU CI Pipeline, Linux GPU TensorRT CI Pipeline, Windows GPU CI Pipeline, Windows GPU TensorRT CI Pipeline, orttraining-amd-gpu-ci-pipeline, orttraining-linux-gpu-ci-pipeline, Linux MIGraphX CI Pipeline, Big Models

[azure-pipelines]

Azure Pipelines successfully started running 8 pipeline(s).

[mtavenrath]

When doing this change I had CUDA warnings disabled. Moving to /W4 without CUDA warnigns disabled opens a lot of new warnings, especially in CUTLASS. Closing for now until warning related issues are solved in CUTLASS:

[skottmckay]

Could we call set_msvc_c_cpp_compiler_warning_level from onnxruntime_providers_cuda.cmake to set to 3 and the start and back to 4 at the end to minimize the code compiled with /W3? maybe we could just do that around the include(cutlass).