refactor: leverage MCore TransformerConfig.__post_init__ validation for config fields set in _apply_performance_config

[terrykong]

Context

When _apply_performance_config() (and similar _apply_* functions in nemo_rl/models/megatron/setup.py) set fields on model_cfg, they do so via plain attribute assignment after the TransformerConfig dataclass has already been constructed and __post_init__ has run. This means MCore's built-in validation (e.g., checking recompute_modules against the allowed set) is bypassed for user-supplied values.

Call chain

1. MegatronPolicyWorker.__init__ → validate_and_set_config()
2. validate_and_set_config() → setup_model_config()
3. setup_model_config() calls ConfigContainer.from_yaml() — __post_init__ runs here
4. _apply_performance_config() mutates model_cfg via attribute assignment — __post_init__ does not re-run

Problem

A typo in a user config (e.g., recompute_modules: ["mopee"]) is silently accepted. MCore's runtime checks match module names via string comparison, so an unrecognized name is a silent no-op — the user thinks they're saving memory but nothing is actually recomputed.

Proposal

Consider refactoring the config initialization flow so that MCore's __post_init__ validation runs after all _apply_* mutations are complete. Options include:

• Constructing the TransformerConfig from a merged dict (user overrides applied first, then construct once)
• Re-running __post_init__() after all mutations (need to verify safety — it also sets defaults)
• Extracting MCore's validation into a callable utility and invoking it post-mutation

This would eliminate the need to duplicate validation logic in NeMo-RL and ensure any future MCore validation additions are automatically picked up.

Related

Surfaced during review of PR #2280 (selective activation checkpointing).