[rtl] Refactor mem_downsizer for full throughput + min latency

[thomas6785]

mem_downsizer no longer pipelines requests, allowing minimal latency.
Downstream transactions are are now pipelined to allow full throughput.
An optional parameter CUT() can be set to add a pipeline stage if timing closure becomes an issue

This reduces ROM read latency from 10 to 6 cycles and increases maximum throughput from 2 to 4 bytes per cycle

Some reservations I hold about this:

• The design is split into some submodules which I've put in the same file - I don't see much value in creating separate files for a single-purpose module, but Verilator dislikes it. Hoping someone can advise if it's really necessary to place these in a separate file; for now I've put in an inline waiver
• I've included the testbench in this commit, which I used as a sanity check. However it's very messy and large sections are AI-generated, so it may be best to leave this without a testbench as it is sufficiently exercised by the toplevel anyway.

Would appreciate if somebody could weigh in on these points then I will remove draft status 😄