add outlier detector, test for it, and start tracking z score of loss

[karpathy]

still TODO:

• track grad norm in addition to loss (do we have to split the gpt2_update function)
• add argparse option to monitor loss and grad norm for outliers (e.g. z > 3)
• add handling of instabilities (e.g. if z > 3, skip update)
• possibly: instead of skipping update we could EXIT_FAILURE, and when -y 1 resumes the optimization from last checkpoint, we could try to somehow advance the DataLoader past the point of instability (this is what PaLM did)

[karpathy]

Introduced two new flags:

    fprintf(stderr, "  -sl <float> outlier stability: skip update if loss goes above this in zscore (0.0f=off, default=3.0f)\n");
    fprintf(stderr, "  -sg <float> outlier stability: skip update if grad_norm goes above this in zscore (0.0f=off, default=3.0f)\n");

They default to 0.0 (old behavior) but if they are e.g. 2.5, then if the loss or grad z score are > 2.5, the update is skipped.