Use `llvm.func`'s `reqd_work_group_size` to specify static local size

[victor-eds]

#2770 tried changing max_work_group_size to reqd_work_group_size as this may be beneficial for performance. However, CI failed for some tests (see).

Find out why CI was failing (are we changing the number of warps when lowering?), fix whatever issue and set reqd_work_group_size when possible.

• If it is always possible to set reqd_work_group_size, do not set max_work_group_size.
• If it is never possible to set reqd_work_group_size, find out if using max_work_group_size is fine, i.e., if the value we are providing is an upper bound for the local size. Also, whether it is worth it, i.e., if IGC is using it at all
• If it is sometimes possible, set it when possible, find out if using max_work_group_size is fine when we cannot set reqd_work_group_size, i.e., if the value we are providing is an upper bound for the local size. Also, whether it is worth it, i.e., if IGC is using it at all.

[victor-eds]

Blocked by backend bug

[victor-eds]

Still blocked by backend bug

[sommerlukas]

Assigning to @whitneywhtsang as she's the assignee on #2845.