add interactive mode

[andrewkchan]

Adds an "interactive mode" that loads the model weights and then continuously prompts the user for new commands. Added because loading deepseek weights is truly horrendously slow and sometimes I want to test multiple things at a time.

Use it like so:

# ./main <checkpoint dir> -m i -L
# Load the checkpoint dir in interactive mode with -L to lock the weights in memory
./main v2-lite-f16 -m i

Then at the prompt, input commands similar to how you would input commands in the CLI, except without the checkpoint dir and omitting -m:

> c -i "Q: What is the meaning of life?" -n 16 -t 0.5
# Equivalent to ./main v2-lite-f16 -m c -L -i "Q: What is the meaning of life?" -n 16 -t 0.5