Fix training mode for LiRA code at inference

[dedeswim]

While working on a follow-up work to the LiRA and the Truth Serum papers, I realized that, when doing inference to get the model's logits, the model is called in training mode:

privacy/research/mi_lira_2021/inference.py

Line 72 in 97eec1a

logits = model.model(this_x, training=True)

This issue was first raised in #427.

Some of the authors (@carlini and @srxzr) confirmed to me that this is a mistake.

This PR introduces the following changes:

• Update the way the model is called at inference time
• Update the results to be expected from the demo runs for both the LiRA and the Truth Serum papers