Scalability of TGM

[shenyangHuang]

Description

1. separate backend for tguf, tguf backend
2. dedicated tguf IO
3. uniform sampling won't work with tguf
4. API on DGraph, number of nodes, number of edges etc. changes needed tguf, get it from the tguf meta-data, potentially edit tguf meta-data
5. dataloading: pre-fetch all the code execution, add pre-fetching to tgm
6. pre-fetching and executing hook
7. optional model sharding

https://github.com/Jacob-Chmura/tgn.cpp

[shenyangHuang]

workflow:

1. the user produces the tguf file from dataset. write a tguf IO
2. tgm reads tguf from disk