Poseidon integration tests fix

[MaksymMalicki]

No description provided.

[MaksymMalicki]

After comparing our memory with rust I've noticed, that in lambdaclass they are only infering the memory values for the segment addresses, which were found during getting op0 and op1. The memory values under those addresses are initially of type unknown. They decided to store all the produced hash inputs and outputs in a separate cache (address - mv map) and then populate only the memory values under the relevant addresses. What we were doing (and still are in different builtins), is that we are immiadetely writing the produced hash to the memory instead of cache, populating also the addresses which are apparently not relevant to the execution. That was the reason of the difference in the traces.

For instance in the screenshot below, I printed only the addresses found during getting op0 and op1 , which should infer the values for.

By writing the entire hash to the memory, we populated the holes between addresses (like 9:3 - 9:9) and thus the differences in the trace.

If you guys agree with the usage of cache, we should implement it for the ec_op and keccak, bc they write to more than one cell.

[TAdev0]

lgtm!