fix: define LAMBDA_ARN_REGEX in finetune_utils to fix RLVRTrainer NameError

[jam-jee]

Issue

RLVRTrainer.train() crashes during reward-function verification with:

NameError: name 'LAMBDA_ARN_REGEX' is not defined
  File ".../sagemaker/train/rlvr_trainer.py", line 280, in _verify_reward_function
    elif isinstance(reward_function, str) and not _is_lambda_arn(reward_function):
  File ".../sagemaker/train/common_utils/finetune_utils.py", line 454, in _is_lambda_arn
    return bool(LAMBDA_ARN_REGEX.match(arn))

_is_lambda_arn() in finetune_utils.py references LAMBDA_ARN_REGEX, but the constant is never defined or imported in that module. Any RLVR training submitted with an evaluator (hub-content) ARN reward function takes the not _is_lambda_arn(...) branch in RLVRTrainer._verify_reward_function and hits the NameError, blocking programmatic RLVR submission entirely.

Fix

The identical regex already exists in rlvr_reward_verifier.py, and both call sites answer the same question — "is this reward-function string a Lambda ARN?". Rather than add a second, drift-prone copy, finetune_utils.py now imports the canonical pattern:

from sagemaker.train.common_utils.rlvr_reward_verifier import LAMBDA_ARN_REGEX

No circular import is introduced (rlvr_reward_verifier does not import finetune_utils).

Tests

Adds TestIsLambdaArn covering:

• valid Lambda ARNs (incl. non-aws partitions like aws-us-gov)
• evaluator hub-content ARNs → not a Lambda ARN
• arbitrary non-ARN strings
• assertion that both modules share the same compiled regex object (guards against future divergence)

Notes

A second, independent bug in the same RLVR verification path (Evaluator.get() raising ValueError on an unrecognized EvaluatorMethod) was identified but is intentionally out of scope for this PR and will be addressed separately.