Small sputnik wins

[matklad]

I think I've seen a handful 1-5% optimizations for sputnic. In isolation, they don't seem to make a lot of sense, but together they should add up. This issue collects the list of things I've tried and found beneficial or which think would be benefitial

These things were measured to be ~1% improvements:

• returning an &H256 from peek, to avoid copy
• refactoring record_dynamic_cost to not contain ? (the whole business of storing Result as a field and checking it on every access feels suspect)
• storing code in Bytes rather than Rc<Vec<u8>> to avoid extra indirection during opcode fetch
• use ArrayVec<U256, 1024> rather than Vec<U256> for the stack

These are theoretical wins

• re-writing pushes to use little endian (trick from evmone https://github.com/ethereum/evmone/blob/d87be127fe15678badc52c3da8f4fa74dc87dd46/lib/evmone/advanced_analysis.cpp#L138-L139). Pushes are common. Also, if we do this in-place, that'd be very cute :-)
• padding the code to avoid branch on push
• merging gas metering and execution
• eliminate stack checks? I think static semantics ensure that VM can never overflow/underflow the operand stack...