[Handshake] Increase LSQ load/store queue to hold largest group

[MaxWipfli]

If a single basic block (= group) has too many loads/stores, the LSQ must use larger load/store queues since allocation needs to happen at once. If there are at most L load and S stores in a single group, then the LSQ's load queue depth must be L+1 and the store queue depth must be S+1.

An assertion to ensure this in the LSQ generator was already added in #993. This now ensures that the C++ code always emits valid load/store queue sizes.

This fixes #832.

[murphe67]

oh this is an interesting idea, yes makes sense 😁

[MaxWipfli]

There is an alternative solution to handle large groups without exploding queue sizes (e.g., if there is a large group that is only executed once per kernel):

In particular, a single group could be split into multiple smaller groups (either in C++, or this could even be done by the LSQ generator internally), so that a single basic block could consist of multiple groups. However, this should only be implemented if it actually becomes a problem in practice.