It's quite common to have a class where we almost always have at most
a single allocated instance (per thread). Now these instances can be
allocated more quickly, by reusing the memory that was used for the
most recently freed instance (a separate memory block is reused for
each thread on free-threaded builds).

It's used like this (only the value 1 is supported for now):
```
from mypy_extensions import mypyc_attr

@mypyc_attr(free_list=1)
class Foo:
    ...
```

This makes a microbenchmark that only allocates and immediately
frees simple objects repeatedly around 3.8x faster.

It's probably worth extending this to support larger free lists in the
future.

We can later look into enabling this automatically for certain native
classes based on profile information.