MMU/MMIO performance improvements for trace production

[OlivierBBB]

Main point

Trace production for the MMU/MMIO is currently highly suboptimal. This is well-known to both @letypequividelespoubelles and @OlivierBBB. Here are 3 ways to handle memory:

• current solution: full snapshots of (the EVM's) memory every time the HUB calls upon the MMU module
• improved solution: partial snapshots of (the EVM's) memory (i.e. capturing all limbs which may be read from / written to
• ambitious solution make no such snapshots, rather replay all memory operations at commit time

Live-tracing instructions is an option, but it would bring complications due to reverted LOG's, e.g. uncertainty about MMU/MMIO stamps. These could likely be handled, and it may represent the best alternative to the current solution.