Make all jumps use a two "word" instruction format.

[markshannon]

Discussed in #297

Using inline caches has some nice advantages in terms of cache locality, and should get us a several percent speedup, but not if it makes all jumps slower due to extra EXTENDED_ARG instructions.

As an initial implementation, we will keep all the original jump instructions (including FOR_ITER) and just make them two "code units" long.
The second oparg be the low 8 bits of the jump. Jumps over 2**16 can use EXTENDED_ARG as before.

The change is simple, all jump instructions get this extra code at the start:

    oparg = (oparg<<8) | ((uint8_t *)next_instr)[1];
    next_instr++;

The compiler will need to account for the addition offset, but that should be simple.

We can then implement a more principled approach later. Ideally, for 3.11 if we have time.

[markshannon]

Postponing this until inline caches are fully implemented. Then we can re-assess what changes would be worthwhile.

[markshannon]

With relative branches, and better basic block layout, it looks like this is unnecessary.