Enable scgpt for cuda-12.8 environment

[lidanwu]

PyTorch 2.7.0 and later (including 2.7.1 and nightly builds) introduce support for the NVIDIA Blackwell architecture (sm_120). This support requires using CUDA 12.8 or 12.9, with earlier PyTorch versions (<=2.6) not supporting these GPUs. This PR enables scgpt inference and training in cuda 12.8 environment that using Blackwell GPU.

• up version to 0.2.5
• support flash-attn 2.8.3 while keep backwards compatibility with flash-attn 1.0.4
• add pure-python Vocab since the development of torchtext has stopped with its latest version at 0.18.0 and torchtext >=0.14 ships Vocab as a C++ extension type that cannot be subclassed like before.
• add more tests and correct legacy issues.

All pytest passed in cuda-12.8 environment that uses flash-attn 2.8.3 but no torchtext

[lidanwu]

I didn't mean to PR into the official repository. sorry for the confusion.