Feat(recursion): tower verifier

[kunxian-xia]

Summary

This is a follow-up work of #1175.

We don't need to write next_layer_evals to the three point_and_evaluations array in each layer. Therefore we can get rid of the cost of two for-loops.

1. builder.range(0, num_prod_spec).for_each();
2. builder.range(0, num_logup_spec).for_each().

Performance impact

With 10k keccak invocations, the number of cycles spent in Ecall_Keccak's tower verifier is reduced.

	shard 0	shard 1	shard 2
before	423951	737341	707351
after	277885 (~140k, 35%⬇️)	323029 (~410k, 56%⬇️)	317425 (~390k, 55%⬇️)

The overall cycles looks like this.

	shard 0	shard 1	shard 2
before	3908867	5008107	5174633
after	3749371 (3%⬇️)	4583889 (8%⬇️)	4768503 (8%⬇️)