fix metrics calculation

[sadrasafa]

The original code calculates the metrics incorrectly.
PSNR (Mathematical Correction): The original code computed a global average MSE across the entire 5D video tensor (B, C, T, H, W) before converting to log space. This made the evaluation strictly batch-size dependent. Furthermore, because logarithms are non-linear, averaging linear MSE values yields a mathematically different result than averaging per-sample decibel values. The fix computes the MSE per sample, converts each to PSNR individually, and then averages the resulting values.

MS_SSIM (Shape Mismatch Fix): The original code passed a 5D video tensor (B, C, T, H, W) to ms_ssim, which expects a standard 4D image batch (N, C, H, W). Flattening the time dimension into the batch dimension resolves this structural bug.

LPIPS (Redundant Normalization): Eliminated unnecessary computation where the data was scaled to [0, 1] only to be mapped back to [-1, 1] for the LPIPS forward pass.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.