Paged attention new

[Bob-Chen222]

Description of changes:

I have added page attention for the spec scheduler. I will clean up the print statement and add more documentation tomorrow. Let me know if anything needs to be changed!

One thing to notice is that both the specscheduler branch and this branch suffer from an "invalid argument" error in Cuda, but I think a small fix would solve this problem on both branches.

---

This change is 

[Bob-Chen222]

Some more updates:

1. added max-kv-cache-size as a flag. At initialization, the page manager will be initialized with the number of hidden layers as input so that the page manager will know how much kv cache will have to be allocated per transformer layer. Then when we init the metadata of inference operators, we call page manager to get the kv cache size needed and allocate the slots accordingly.
2. add paged attention support for incr_decoding
3. clean up the comments and reorganize the format.
4. after all these changes the performance is nearly the same as before

[lockshaw]

Moved to flexflow/flexflow-serve#82