Fix TimesFM patch normalization instability

[AnMakc]

This PR fixes patch normalization issue in TimesFM model:

• Numerical instability for low-variance data
• Incorrect clamping of calculated sigma to one instead of configured eps

Issue only affects data with low relative variance and results in incorrect prediction.
Variance is now calculated after mean removal to avoid catastrophic cancellation.

@kashif Please, take a look as you contributed this model.

Bug demo

Code to reproduce

import random

import numpy as np
import matplotlib.pyplot as plt
import torch

from transformers import TimesFmModelForPrediction


seed = 42
np.random.seed(seed)
torch.manual_seed(seed)
random.seed(seed)


model = TimesFmModelForPrediction.from_pretrained(
    "google/timesfm-2.0-500m-pytorch",
    attn_implementation="sdpa",
    device_map="auto",
)

model.eval()


X_inp = np.arange(256)
X_pred = np.arange(256, 256 + 128)

fig, axs = plt.subplots(2, 2, figsize=(12, 12))

for ax, sigma in zip(axs.flatten(), [1e-0, 1e-1, 1e-2, 1e-3]):
    rng = np.random.default_rng(seed=seed)
    Y_inp = 1e3 + rng.normal(loc=0, scale=sigma, size=(256,)).cumsum()

    inputs = [
        torch.from_numpy(Y_inp).to(model.device, dtype=model.dtype)
    ]

    with torch.no_grad():
        pred = model(past_values=inputs)
        pred = pred.mean_predictions.cpu().numpy().squeeze()

    ax.plot(X_inp, Y_inp, label='Inputs')
    ax.plot(X_pred, pred, label='Mean Prediction')
    ax.set_title(f'Sigma={sigma}')
    ax.legend()

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: timesfm