Fixes #46.

Benchmarks show a dramatic improvement; not quite as good as the
demonstration in #46 because the implementation had to be a bit more
complex to properly handle unregistration, but still very good. Here it
is with 20,000 items already registered:
```
   %time add_to_reg(reg, 1000)
   CPU times: user 190 ms, sys: 19.3 ms, total: 209 ms
   Wall time: 203 ms
```

Here's a profile with about 100,000 utilities registered:
```
 %prun add_to_reg(reg, 1000)
         80005 function calls (79005 primitive calls) in 0.713 seconds

   Ordered by: internal time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
     1000    0.596    0.001    0.621    0.001 adapter.py:202(subscribe)
    12000    0.016    0.000    0.019    0.000 interface.py:518(__hash__)
     1000    0.010    0.000    0.709    0.001 registry.py:206(registerUtility)
3000/2000    0.009    0.000    0.014    0.000 interface.py:255(interfaces)
     2000    0.008    0.000    0.022    0.000 adapter.py:637(changed)
     1000    0.008    0.000    0.025    0.000 registry.py:498(_getUtilityProvided)
     9000    0.008    0.000    0.017    0.000 {method 'get' of 'dict' objects}
     1000    0.008    0.000    0.027    0.000 adapter.py:102(register)
     2000    0.006    0.000    0.008    0.000 adapter.py:450(changed)
     1000    0.005    0.000    0.663    0.001 registry.py:145(registerUtility)
```

I was very careful not to change the pickle at all.

zope.interface and zope.component tests have been run and both pass. (It
was necessary to account for the underlying objects changing because of
the way zope.component cleans up after tests.)

We exclude interfaces declared with the 'class' statement, because their
methods typically have only a doctstring in the suite, which confuses
the branch coverage machinery.