MathInterpreter: Replace Instruction polymorphism with bytecode

[mspraggs]

Hi Antonin,

I hope you're keeping well.

This change replaces the Instruction class hierarchy with a bytecode-based instruction representation. Instead of an Instruction instance, each instruction opcode is specified by a single byte, and instruction arguments are packed into the bytecode array. The arithmetic operations are represented by a single byte, the load, store and call operations are encoded as a byte followed by the variable or function address, and the constant instruction is encoded as a single opcode followed by the double-precision floating point value. I hope this makes sense. If not, this should provide a fuller explanation.

Replacing the polymorphic instruction classes with a packed bytecode array increases the locality of reference during execution. After some rough benchmarks I found the improvement was a factor of two, though more performance could be squeezed out through other modifications. For example, the std::deque container wrapped by the std::stack could be replaced with a std::vector.

Let me know what you think.

Best,

Matt