feat: environments folder

[cmunley1]

See #986

Introduces environments/ folder

Separates environment configuration (agent + dataset pairings) from server implementation configuration to improve environment discoverability and organization.

Added a top-level environments/ directory with three subdirectories:

• environments/training/ 47 training environments (one per dataset)
• environments/eval/ 10 evaluation benchmarks (replaces benchmarks/)
• environments/example/ 14 simple examples configs

Total: 73 environments

Config separation

• resources_servers/*/configs/*.yaml now contain server parameters only
• Dataset definitions and agent pairings have been moved to environments/

Each environment config:

• includes a config_paths entry pointing back to its resources server config
• defines the agent + dataset section

Motivation

Previously, each YAML file in resources_servers/*/configs/ served two purpose:

• server implementation config
• dataset registry

This created limitations on environment discoverability:

• Could not define multiple training environments using the same resources server
  • e.g. dapo_math_17k and open_math_reasoning both depend on math_with_judge

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.