Possible solution of using DLPack to handle TVM FFI

[haoran35-jpg]

1. Hey! I was trying to research for opportunity of lowering JAX to ThunderKittens, and met the same issue as mentioned in [TVM FFI support #176]. I searched internet and figured that DLPack kind of borrows the idea of TVM FFI, although there is certain differences: https://tvm.apache.org/ffi/concepts/tensor.html. I tried to implement DLPack methods and tested with both routes, getting the following difference in VRAM use:

   Without DLPack	Using DLPack
   100.000 MB	50.000 MB

The two routes are respectively: tensor.clone() -> TKParallelTensor(tensor, ...) and JAX → __dlpack__() → TKParallelTensor(capsule). The VRAM use is integer because I'm only testing with a single fixed tensor size (50 MB)—i.e. 50×1024×1024/4 float32 elements—so the copy path allocates exactly two buffers (100 MB) and the DLPack path one (50 MB), giving round numbers. Not sure if this is in a good shape of PR? Thank you!