Add CastStar GIFT-Eval results (#160)

CastStar is an agentic time-series forecasting system built on top of
multiple foundation forecasting models. The submitted GIFT-Eval result
uses a meta-selection/ensemble layer over a foundation-model pool and is
intended for leaderboard evaluation under the GIFT-Eval protocol.

Model type: agentic
Organization: USTC-AGI
Test-data leakage: No
Replication code available: No
97/97 dataset configurations, 11 metrics, official CSV format

Model pool:
- Chronos-2
- TimesFM-2.5
- FlowState
- TiRex
- PatchTST-FM
- Toto-2.0-2.5B

Files added:
- results/CastStar/all_results.csv
- results/CastStar/config.json

Model card: https://huggingface.co/CastStar/CastStar
Code link: https://github.com/ustc-time-series/CastStar

Co-authored-by: fishsure <fishsure@users.noreply.github.com>

Author: fishsure