FlashInfer Windows support

[SystemPanic]

This PR adds support for Windows OS, including AOT and JIT builds.

Due to standard console (cmd.exe) command length limit of 8192 chars, PowerShell is required to do the installation / build.

A Visual Studio 2019 or newer is required to launch the compiler x64 environment. The installation path is referred in the instructions as VISUAL_STUDIO_INSTALL_PATH, just replace it with your Visual Studio installation path.

Instructions for Windows build:

1. Open PowerShell
2. Clone the repository
3. Change the working directory to the cloned repository path, for example: cd C:\flashinfer
4. Execute the following commands in order:

Import-Module "VISUAL_STUDIO_INSTALL_PATH\Common7\Tools\Microsoft.VisualStudio.DevShell.dll"
Enter-VsDevShell -VsInstallPath "VISUAL_STUDIO_INSTALL_PATH" -DevCmdArguments '-arch=x64'
$env:DISTUTILS_USE_SDK=1;
$env:FLASHINFER_ENABLE_AOT=1;
$env:MAX_JOBS=10;

5. (Optional) To speed up compilation time, set environment var TORCH_CUDA_ARCH_LIST to match your gpu arch. For example: $env:TORCH_CUDA_ARCH_LIST="8.6";
6. (Optional) To use the installed torch in the case you have a nightly / built from source version, execute: $env:FLASHINFER_USE_CURRENT_TORCH=1;
7. Execute: pip install . --no-build-isolation to install or python setup.py bdist_wheel to generate a installable Windows wheel
8. (Optional) Clean the Windows customized build folder: python setup.py clean

Special mention to changes made on:

• pod.cu: To solve a C1060 compiler error, the PrefillParams and DecodeParams assignments has been moved to an external function.
• setup.py: To support current and future compiled kernels object linking, the build directory has been customized on Windows to keep the final linker command below the PowerShell limit of 32767 characters,