Replace libcudart.so with PyTorch's CUDA APIs

[rapsealk]

This PR resolves #264
I opened this PR as #346 has been closed accidently.

[atkinson]

Pretty please, can this be merged and released?

[rapsealk]

@atkinson Glad to hear that it was helpful to you. I don't have sufficient authority to merge it by myself, so let's wait for @TimDettmers' feedback.

[Titus-von-Koeller]

@rapsealk Thanks for your input, this is very much appreciated! I looked into this, talked it through with Tim and will work on integrating this later this week. Basically, we will take your solution and combine it with some other tweaks that we have in mind to improve this module.

Will let you know once we have news.

[rapsealk]

@Titus-von-Koeller Thanks for the news! It would be a pleasure to help if needed.

[TimDettmers]

Thank you for the PR. This will resolve a couple of problems where the libcuda.so cannot be found. There are other issue the PR might create. Sometimes a user needs to use a different CUDA version for pytorch and bitsandbytes, but this is currently rare.

I am unsure how this fix behaves if another library version loads a different CUDA version before PyTorch loads its CUDA version. Since PyTorch is packaged with CUDA its not a problem for them, but it might create problems for bitsandbytes.

I will merge and try to write some test to check behavior with multiple CUDA versions installed. Otherwise, we can just hope it works for others with weird environments.

Thanks again for your contribution!