Stream logs unified improvement

[amazeAmazing]

Issue #, if available:

Description of changes:

fix(stream_logs): unify log streaming for MTRL, evaluators, and HyperPod

Summary

• MTRL stream_logs() now uses correct log group (/aws/sagemaker/Job/AgentRFT) and polls via Job API — previously hung forever
• SMTJ _stream_logs_smtj() refactored to use LogStreamer + stream_log_loop — same timeout/feedback behavior as MTRL and evaluators
• Evaluators get stream_logs() on BaseEvaluator (pipeline and HyperPod dispatch)
• HyperPod _stream_logs_smhp() provides user feedback instead of silent failure
• Notebook examples added for all evaluator types + MTRL trainer

What changed

File	Change
common_utils/log_streamer.py	New. LogStreamer (poll-once CW utility) + stream_log_loop() shared helper. Warns after ~30s of no events, exits after ~5min.
agent_rft_job.py	Add stream_logs()
multi_turn_rl_trainer.py	Override stream_logs() with correct log group
base_trainer.py	Refactor _stream_logs_smtj() to use LogStreamer + stream_log_loop. Patch _stream_logs_smhp() with ClientError handling + feedback.
evaluate/base_evaluator.py	Add _latest_execution (set by _start_execution and _submit_hyperpod_eval_job) + stream_logs() with HP/pipeline dispatch
Notebooks (5 files)	Add # evaluator.stream_logs() / # job.stream_logs() examples

Testing

Unit tests (35):

• test_log_streamer.py: LogStreamer stream/filter mode, dedup, error handling, stream_log_loop early-exit/terminal/feedback/retry/timeout
• test_stream_logs.py: AgentRFTJob, MultiTurnRLTrainer, BaseEvaluator, ARN resolution

Integration tests (8, no new jobs launched):

• test_stream_logs_trainer.py (us-west-2 / 729646638167): MTRL via AgentRFTJob.get(), serverful SMTJ, input validation
• test_stream_logs_evaluator.py (us-west-2 / 729646638167): BenchMarkEvaluator, CustomScorerEvaluator, LLMAsJudgeEvaluator — each with real completed pipeline execution ARNs

Open questions

• AgentRFTJob.stream_logs() exists for the attach-to-existing-job use case (AgentRFTJob.get("name").stream_logs()). SMTJ trainers (SFT/DPO/RLVR) have no equivalent attach-and-stream pattern — users must construct a trainer and stuff _latest_training_job. Should we add a similar wrapper or a standalone helper for SMTJ jobs?

How to verify

# MTRL (was: hung forever)
job = AgentRFTJob.get("my-job")
job.stream_logs()

# Evaluator (was: AttributeError)
execution = evaluator.evaluate()
evaluator.stream_logs()

# HyperPod (was: silent)
trainer.stream_logs()  # now shows "Waiting for logs..." feedback

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.