Add pretrained weights

[grey-nearing]

Add pre-trained model weights

Introduce two sets of pre-trained model runs (Mean Embedding Forecast
LSTM) based on the Google Flood Hub architecture to support transfer
learning, local fine-tuning, and Prediction in Ungauged Basins (PUB).

Models included:

• google-floodhub-base (google-floodhub-settings-55-epochs): Full basin baseline capturing global
  hydrological variety.
• google-floodhub-nse-filtered ('google-floodhub-settings-55-epochs-nse-filtered-0.5-85-epochs'): High-skill baseline filtered
  for catchments with historical NSE > 0.5.

Each model runtime release includes weights (.pt), optimizer states,
pre-computed global data scalers, and the exact configuration files
for full reproducibility.

These models were trained on the FULL historical data period (1982-2023)
with NO temporal holdout/test split. Evaluation on this historical timeline
will yield fundamentally invalid, artificially inflated metrics due to
in-sample data leakage.

Appropriate use cases include:

• Local fine-tuning / transfer learning (with lower learning rates)
• Spatial generalization (PUB) on entirely excluded basins
• Forward-looking inference on post-2023 data periods

[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.

[PatrickGoogle]

LGTM