Feature Request: Support for multiple validation sets + separate loss validation reporting

Currently, Megatron-Bridge supports evaluating on a single validation set. For many workflows (e.g., multi-domain SFT, benchmarking across datasets), it is useful to:

1. Define multiple validation datasets in the config.
2. Report losses separately for each validation set (instead of aggregating).

This would allow fine-grained monitoring of model performance across domains or tasks without maintaining separate training runs.

The issue does not name files or tests. Start by locating the configuration for validation datasets and the evaluation and loss-reporting entry points, then trace how the current single validation set is defined and evaluated. Done means multiple validation datasets can be configured and losses are reported separately for each set, with coverage for both behaviors.