feat: add aws-rft-sdk package for Reinforcement Fine-Tuning

[xiaoxshe]

Summary

• Add standalone aws-rft-sdk package that integrates SageMaker RFT (Reinforcement Fine-Tuning) with Strands agent framework
• RolloutFeedbackClient: calls real CompleteTrajectory and UpdateReward APIs via SigV4-signed requests
• @rft_handler: decorator to extract rollout metadata from payloads and manage RFT context
• RFTContext: thread-local context that produces X-Rft-Job-Arn, X-Trajectory-Id, X-Span-Id headers
• wrap_model: Strands model adapter that injects these headers into every inference call via extra_headers

API Integration

Uses the AgenticRFTRuntimeService alpha endpoint (finetuning-job-runtime.alpha.sagemaker.us-west-2.api.aws):

• POST /CompleteTrajectory — marks trajectory as complete (PENDING -> READY)
• POST /UpdateReward — submits per-transition rewards (READY -> REWARD_RECEIVED)
• POST /sample — OpenAI-compatible inference proxy with trajectory capture via X-Rft-Job-Arn, X-Trajectory-Id, X-Span-Id headers
• SigV4 signing with service name sagemaker

Test Results

Against alpha endpoint:

• CompleteTrajectory: 404 for non-existent trajectory (API reachable, auth working)
• UpdateReward: 404 for non-existent trajectory (API reachable, auth working)
• SampleInvoke: 400/500 with validation errors (API reachable, requires valid job ARN)

Unit/Integration tests (5/5 pass):

• RFTContext produces correct X-Rft-Job-Arn, X-Trajectory-Id, X-Span-Id headers
• wrap_model injects headers into inner model's stream() call
• RolloutFeedbackClient sends SigV4-signed CompleteTrajectory request
• RolloutFeedbackClient sends SigV4-signed UpdateReward request
• End-to-end: @rft_handler -> RFTContext -> wrap_model -> correct headers injected

Strands SDK compatibility:

• 72/72 OpenAI model tests pass (no regressions from extra_headers change)