ci(fp8): fix FP8 precision tests

[CyCle1024]

Summary

• Add replay-based k-grouped GEMM FP8 coverage against a FP32 scaled-grouped-mm reference.
• Enable deterministic algorithms in testing utilities and tighten FP8 MoE train-engine loss references.
• Update the unit-test CI image to the current FP8-capable image.

Tests

• Not run; PR opened from the existing branch commits.