bpo-42246: Eliminate jumps to exit blocks by copying those blocks.

[markshannon]

More yak-shaving for PEP 626.

This PR:

• Marks all blocks that exit the scope or that do not fall through to the b_next block.
• Removes the a_reverse_postorder array in the assembler, as it is just a copy of the b->b_next chain.
• Eliminates unconditional jumps to short exit blocks, by copying the exit blocks onto the end of its predecessor.

The net effect of this is that:

def f(c):
    if c:
        A
    else:
        B

now compiles to:

  2           0 LOAD_FAST                0 (c)
              2 POP_JUMP_IF_FALSE       12

  3           4 LOAD_GLOBAL              0 (A)
              6 POP_TOP
              8 LOAD_CONST               0 (None)
             10 RETURN_VALUE

  5     >>   12 LOAD_GLOBAL              1 (B)
             14 POP_TOP
             16 LOAD_CONST               0 (None)
             18 RETURN_VALUE

instead of

  2           0 LOAD_FAST                0 (c)
              2 POP_JUMP_IF_FALSE       10

  3           4 LOAD_GLOBAL              0 (A)
              6 POP_TOP
              8 JUMP_FORWARD             4 (to 14)

  5     >>   10 LOAD_GLOBAL              1 (B)
             12 POP_TOP
        >>   14 LOAD_CONST               0 (None)
             16 RETURN_VALUE

This has two effects

1. Means that f_lineno can be determined from f_lasti even after an implicit return. This is needed for PEP 626.
2. Reduces the number of bytecodes executed at the cost of a minor increase in bytecode size, and might result in a tiny speedup. I haven't measured it, as I expect the change to be too small to meaningful, or even measurable.

https://bugs.python.org/issue42246

[bedevere-bot]

@markshannon: Please replace # with GH- in the commit message next time. Thanks!