feat(recursion): reduce cycles in ceno proof loading

[kunxian-xia]

Summary

• rely on Feat: read prod_evals and logup_evals from hint in sumcheck_layer_eval openvm#30.
• we read prod_evals from hint space inside NativeSumcheck chip instead of moving them from hint stream to memory word by word.

Performance

verifying the 6th shard of block 23817600, the total cells and instructions_executed before and after this PR.

	total cells	instructions_executed
before	2B	9.5M
after	1.3B	6.7M

[hero78119]

LGTM 👍

some other notes per offline discussion

1. constrain consecutive evals between sumcheck_layer_eval
2. add evals to transcript, follow #1125