Which would've saved me lots of time and headbanging earlier.

Each VM doesn't need its own opcode array (hopefully); use a
pre-allocated memory bank for the memory of all amplifiers.

Specify whether the VM uses IO or not, and if not don't create
channels. These are more expensive than I expected, and
pointlessly creating/destroying them thousands of times is
a significant CPU drain.

- Store opcodes in an array rather than a map for faster lookup
  (this means storing them as interface{} as otherwise the types
  form a circular dependency, which is icky but worth the
  performance boost).
- Don't dynamically create an 'Arg' function every step of the vm.
  This was *super* expensive.
- Calculate a bitmask of the argument modes instead of relying
  on on-demand div-modding.
- Only buffer a single value in the I/O buffers to reduce the
  size of allocations.

For the most part we run for a few milliseconds and quit, there's
no need to do any GC in that time.

This will almost certainly bite me in the ass later somehow, but
for now it gives a small speedup.