[Feature Request] Better load/build inline

[DarkSharpness]

Great work! I have a small suggestion. Currently, tvm_ffi.cpp.build_inline only accepts raw cpp/CUDA source code strings, rather than file paths. It would be useful to add a tvm_ffi.cpp.load function, similar to torch.utils.cpp_extension.load, which can directly compile from source files.

[junrushao]

Seems like a good first issue! @DarkSharpness feel free to contribute if you have bandwidth, and I'm happy to help if you got any issues/questions.

Basically we need to extend this file: python/tvm_ffi/cpp/load_inline.py, and support Path-like objects.

[yzh119]

FYI, in flashinfer the JitSpec basically implement this functionality on top of tvm-ffi: https://github.com/flashinfer-ai/flashinfer/blob/main/flashinfer/jit/core.py#L180-L280 (with functions such as build (for AOT) and build_and_load (for JIT)).