Explain that unhandled error case is very unlikely

for new allocations, but not when a memory block was already resized

Apparently, the number of objects with docstrings in builtins varies
with --with-pydebug (non-debug has one fewer).

Also, skip the new tests entirely if built --without-doc-strings.

written in C.

As a part of this, a few doctests have been added to the builtins module
(on hex(), oct(), and bin()), a doctest has been fixed (hopefully on all
platforms) on float, and test_builtins now runs doctests in builtins.

This will allow us to control in the future whether to use framing or not.
For example, we may want to turn it off for tiny pickle where it doesn't
help.

The change also improves performance slightly:

### fastpickle ###
Min: 0.608517 -> 0.557358: 1.09x faster
Avg: 0.798892 -> 0.694738: 1.15x faster
Significant (t=3.45)
Stddev: 0.17145 -> 0.12704: 1.3496x smaller
Timeline: http://goo.gl/3xQE1J

### pickle_dict ###
Min: 0.669920 -> 0.615271: 1.09x faster
Avg: 0.733633 -> 0.645058: 1.14x faster
Significant (t=5.05)
Stddev: 0.12041 -> 0.02961: 4.0662x smaller
Timeline: http://goo.gl/LpLSXI

### pickle_list ###
Min: 0.397583 -> 0.368112: 1.08x faster
Avg: 0.412784 -> 0.397223: 1.04x faster
Significant (t=2.78)
Stddev: 0.01518 -> 0.03653: 2.4068x larger
Timeline: http://goo.gl/v39E59

### unpickle_list ###
Min: 0.692935 -> 0.594870: 1.16x faster
Avg: 0.730012 -> 0.628395: 1.16x faster
Significant (t=17.76)
Stddev: 0.02720 -> 0.02995: 1.1012x larger
Timeline: http://goo.gl/2P9AEt

The following not significant results are hidden, use -v to show them:
fastunpickle.

tests (issue13477).

Original patch by Berker Peksag.

for some builtins.

over all CPython source files.

"format units" as legacy converters, as well as two new features:
"self converters" and the "version" directive.