fix: use total denoising steps for Humaneval dual cache evaluation

[Ashleyzm]

This is also inconsistent with the corresponding dual-cache configuration in eval_gsm8k.sh, which passes steps=${length}. For length=256 and block_length=32, the current Humaneval command passes steps=8. Since generate_with_dual_cache computes steps_per_block = steps // num_blocks, this reduces to 1 step per block. This may also address the behavior reported in #66, where dual-cache evaluation on HumanEval produced 0 accuracy.