This PR improves usability significantly for researchers experimenting on streaming generation.

Summary :

in simply/model_lib.py

StreamingToken

• New dataclass for yielded tokens (lines ~3238-3264)

decode_one_step()

• New function for single-step decoding without while_loop (lines ~3937-4039)
  LMInterface.generate_stream() - New generator method that yields tokens as they're generated (lines ~3688-3913)

in simply/model_lib_test.py

• Added test_lm_interface_generate_stream() test case

Usage :

from simply import model_lib

lm = model_lib.LMInterface(model, params, vocab=vocab)

for token in lm.generate_stream("Once upon a time"):
    print(token.token_text, end="", flush=True)
    if token.is_final:
        print()  # newline at end