Fix metrics assertions for Tinker-API based E2E CI runs

[SumanthRH]

Summary

We introduced E2E CI for the Tinker API with the SkyRLTrainBackend in #1616 , but currently we don't have any assertions on the metric values:

SkyRL/tests/train/gpu_e2e_test/gsm8k_tinker.sh

Lines 11 to 13 in 29f0fdf

	# TODO: tighten thresholds after 3-5 nightly runs (5% allowance from min observed),
	# matching the convention in gsm8k_colocate.sh.
	REWARD_MIN_VALUE=0.0

We need to add assertions for expected metrics based on the results of 5 nightly runs