Your issue comes from the breaking change in fastai v1.0.53 making the hidden size a multiple of 8 (1152 instead of 1150) and your pretrained weights have the old size (1150). Just pass along this config to deal with it:
config = awd_lstm_lm_config.copy()
config['n_hid'] = 1150
learn = language_model_learner(data_lm, AWD_LSTM, config=config,
pretrained_fnames=[FILE_LM_ENCODER, FILE_ITOS],
drop_mult=0.3)