Try HPy instead of CPython C-API

[skirpichev]

See https://hpyproject.org/. That should improve speed on PyPy and (especially) GraalPy. See benchmark results for other API for extensions: gmpy2/gmpy2#523 Suggested in pypy/pypy#5147 (comment)

This needs benchmarking on some toy numeric type.

But see #195

[skirpichev]

I did micro-benchmarking for simple type with nb_add slot:

Benchmark	static	heap	hpy
xxx(123) + xxx(321)	202 ns	353 ns: 1.75x slower	276 ns: 1.37x slower

Implementations:

• static type, https://gist.github.com/skirpichev/455d1c3e9484e1d648d48be960feb90b
• heap type, https://gist.github.com/skirpichev/071b8d401678a2ea46f11b34d8ab0d0b
• with HPy, https://gist.github.com/skirpichev/14710326d3fbaab8bdb1d610cff9d838

$ python -V
Python 3.11.12
$ python -c 'import sysconfig;print(sysconfig.get_config_var("CONFIG_ARGS"))'
'--enable-optimizations' '--with-system-libmpdec'
$ echo $CFLAGS 
-Wall -Werror -std=c17 -O3

benchmark code

# bench.py
import pyperf
from operator import add
from example import xxx

x, y = map(xxx, [123, 321])
runner = pyperf.Runner()
s = repr(x) + " + " + repr(y)
runner.bench_func(s, add, x, y)

Note, that HPy currently lacks API for module state (hpyproject/hpy#286). Maybe that's reason for a better numbers c.f. heap type with CPython API.

So, we might expect performance degradation for small arguments, just like in #195. That's not acceptable.