feat: add torch compile

[JINO-ROHIT]

1. adding torch compile degrades performance, possible reasons could be loops and modules that cannot be compiled.
2. try to compile at the leaf node level instead of applying at the top level model

[JINO-ROHIT]

1. we have a lot of graphs being compiled seperately and lots of recompilation, major modules like rope, for loops, kv cache etc is slowing this down a lot.
2. lots of modules needs to be improved to get maximum performance.
3. dynamic shapes and sizes also cause recompilation
4. try using different modes of compilation
5. do this torwards the end when all techniques have been implemented.