fix(serve): BedrockModelBuilder accepts BaseTrainer as model input

[amazeAmazing]

BedrockModelBuilder(model=trainer) now works for SFTTrainer, DPOTrainer, RLVRTrainer, etc. Previously only ModelTrainer and S3 URIs were supported.

Issue #, if available:

Description of changes:

fix(serve): BedrockModelBuilder accepts BaseTrainer as model input

Problem

BedrockModelBuilder(model=trainer) failed when trainer was an SFTTrainer,
DPOTrainer, RLVRTrainer, or any BaseTrainer subclass. Users had to extract the
S3 checkpoint URI manually:

# Worked
builder = BedrockModelBuilder(model="s3://bucket/checkpoint/step_10")

# Didn't work
builder = BedrockModelBuilder(model=trainer)

Changes

• Added BaseTrainer branch in _fetch_model_package() to resolve
  output_model_package_arn from _latest_training_job
• Added manifest fallback in _get_s3_artifacts() for HyperPod jobs where
  model_artifacts is not populated
• Updated type annotation and docstrings to include BaseTrainer

Testing

Unit tests:

Integration tests by resolution path:

Category	Model	Compute	Trainer	Resolution path
Nova (SMTJ)	nova-micro (DPO)	TrainingJobCompute	DPOTrainer	model_artifacts (S3 direct)
Nova (HyperPod)	nova-lite-v2 (RLVR)	HyperPodCompute	RLVRTrainer	manifest fallback
Nova (Bedrock deploy)	nova-lite-v2 (RLVR)	—	RLVRTrainer	manifest fallback → create_custom_model
OSS (SMTJ)	gpt-oss-20b (SFT)	TrainingJobCompute	SFTTrainer	model_package → hf_merged probe
OSS (TrainingJob direct)	gpt-oss-20b (SFT)	TrainingJobCompute	—	model_artifacts (S3 direct)

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