Is your feature request related to a problem? Please describe.
The LithOS report that they rewrite the CUDA kernel functions during runtime, so that they can achieve fine-grained GPU TPC sharing. Can bpftime be used to realize this feature?

Describe the solution you'd like
Could you please provide a sample for this feature?
If bpftime has not supported this, I guess A new bpf program type on GPU, which looks like XDP to net packets, would be helpful to this end.

Describe alternatives you've considered
LithOS did not report their detailed method for code injection, guess it would be tools like nvbit...

Provide usage examples
N/A

Additional context
N/A