ROBs in Streamer for MemPool integration

[mbertuletti]

This PR adds to RedMulE's streaming the reordering logic needed for the integration in MemPool.
Additionally a tiler feature is added, to start computing from any column in W and loop-back once over.

[FrancescoConti]

Ciao @mbertuletti thanks for the PR! I see it already passes CI tests flawlessly, and code-wise I have no particular comment. I am interested in understanding a bit better the loopback feature however, it's not super clear to me.

[mbertuletti]

Ciao @mbertuletti thanks for the PR! I see it already passes CI tests flawlessly, and code-wise I have no particular comment. I am interested in understanding a bit better the loopback feature however, it's not super clear to me.

Ciao @FrancescoConti, sure! I give you a motivation and overview. In TensorPool multiple RedMulE can fetch at the same time from the TCDM. According to the parallelization scheme W is shared, meaning that the RedMulEs will all start fetching the same column of W, creating contentions in the interconnect.

Therefore, I added the option to specify that you want to compute starting from another column (let's say index i out of P total columns). Then you need to compute up to column P-1 and "loopback" to column 0 to complete the computation up to column (i-1). In practice, this is implemented in the Tiler, by adding a loopback state.

The details are in this paper: https://arxiv.org/abs/2604.02291

[FrancescoConti]

Ok - interesting. It makes sense to me. I would add two points (one may actually be already there):

1. is the loopback feature "guarded" by EnableReordering?
2. do you think it is possible to add one or more regression tests specifically for EnableReordering and loopback? Because I fear otherwise we may lose it with subsequent PRs / commits. Of course, if this is too heavy, it may also be done in a future PR (perhaps keeping track of it an issue).