feat: environment suite manager — named multi-env sweeps (M4 core)

[wprazuch]

What

Adds nemo_gym.suite_manager: the M4 core. A suite is a named list of environment names
that expands to the config_paths needed to run them together — so a production sweep over
dozens of environments becomes gym run --suite ultra_v3 instead of a hand-maintained 80+ entry
config_paths.

# nemo_gym/suites/ultra_v3.yaml  (or ~/.config/nemo_gym/suites/<name>.yaml)
description: Ultra V3 evaluation suite
environments:
  - gpqa
  - workplace_assistant

API

• discover_suites() — built-in suites under nemo_gym/suites/ + user suites under
  ~/.config/nemo_gym/suites/; a user suite shadows a built-in of the same name.
• resolve_suite_config_paths(name) — expands a suite into ordered, de-duplicated config_paths
  by resolving each environment name through the registry (feat: environment registry + 'gym list environments' (#1205 friction #8 / M2) #1635). Fast-fails on:
  • unknown suite → SuiteNotFoundError with a "did you mean?" hint,
  • empty suite → EmptySuiteError,
  • unknown environment in a suite → EnvironmentNotFoundError annotated with the suite name.

Like the registry, resolution only reads config files — no interpolation, no servers — so it's
safe even when a referenced config needs unset secrets.

Scope / deferred

• Built-in suites (ultra_v3, reasoning, coding) are intentionally not shipped yet:
  a suite only references environments that exist, and environments/ currently holds one migrated
  env (workplace_assistant). They land here as the environment migration proceeds — see
  nemo_gym/suites/README.md. (No premature suite that would fast-fail on an unmigrated env.)
• CLI wiring (gym list suites, gym run --suite, gym suite add/remove) is deferred to align
  with the unified CLI — same split that worked for the registry (feat: environment registry + 'gym list environments' (#1205 friction #8 / M2) #1635) → run-by-name (feat(cli): run an environment by name — 'gym env run --env <name>' (#1205 friction #8 / FEP-1022) #1636).

Tests

tests/unit_tests/test_suite_manager.py — 13 cases, 100% module coverage: discovery
(built-in/user/shadowing, non-suite & malformed-file skipping, non-string filtering), resolution
(ordering, dedup, all three fast-fail paths), and an end-to-end resolve of the real
workplace_assistant env through a user suite.

Part of the configuration-friction epic (#1205), milestone M4 (friction #8). Targets
wprazuch/environment-registry since it builds on the registry's resolve_environment_config_paths.

[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.