Penalizing system calls and expensive instructions

[fwsGonzo]

Penalize used to reduce max instructions, however it was incompatible with long-running VMs that pause and resume a lot. The reason is that max instructions would eventually become 0 which indicates a stopped machine.

So, for now the solution is to use the regular instruction counter for penalties. It means that instruction counting is now mostly wrong. But perhaps we can one day use a separate penalty counter, where instruction_counter - penalty = real instruction count. The most important thing is to not modify dispatch too much.

[fwsGonzo]

Some work done here: #190

Future work:

• Penalize atomic instructions
•  Penalize expensive bit-manipulation (eg. CLMUL, CLMULR and CLMULH)
• Penalize DIV, DIVU, DIVW, DIVUW, REM, REMU, REMW, REMUW, (?)
• FENCE?