refactor ORT-Extension for the coming GroupQueryAttention work

[jslhcl]

Refactor ORT-Extension for the coming GroupQueryAttention work (https://github.com/microsoft/onnxruntime-extensions/pull/654). Main changes:

• separate common macros to a new header
• separate other ORT dependencies than custom ops to a new header
• cublas and device id info in cuda context
• separate cuda type (MFloat16, etc.) to a new header
• separate cuda device properties to a new header