feat: add max_training_steps config for CI smoke tests

[dinhxuanvu]

Summary

Adds a max_training_steps config field that caps training after N steps regardless of epochs or dataset size. This enables fast CI smoke tests that validate full training pipelines end-to-end without running to completion.

• Supported in all trainer paths: sync RL, async RL, and SFT
• Validated as a positive integer when set (ValueError on <= 0)
• SFT: caps num_steps after resolution from num_epochs and propagates to the LR scheduler
• RL: caps total_training_steps at initialization and adds an early exit after step completion
• Async RL: cancels in-flight generation tasks on early exit to avoid resource leaks

Motivation

Running full training configs in nightly CI is prohibitively expensive. max_training_steps=2 exercises the core pipeline — data loading, model init, forward/backward, and logging — in minutes rather than hours/days, validating that the full stack initializes and trains without errors. Unlike dummy_run_max_steps (SFT-only, uses fabricated data), this uses real data and works across RL and SFT configs.

Test plan

• Unit tests covering config construction, validation, capping logic, early exit, and post-loop cleanup
• 38 tests (33 pass, 5 skip on macOS due to transformers unavailability)
• Pre-commit checks pass (ruff, black)
• GPU integration test with a real config + max_training_steps=2 (requires cluster)

[gemini-code-assist]

Code Review

This pull request introduces a max_training_steps configuration option across RL (sync/async) and SFT trainers, enabling early termination of training regardless of the number of epochs or dataset size. The changes include updates to configuration classes, validation logic, and trainer implementations, along with a new test suite. Review feedback identifies an off-by-one error in the early exit logic of both the synchronous and fully asynchronous trainers, where global_step is incremented before the limit check; this could lead to incorrect checkpointing and skipped steps upon resumption.