add sys.meta_path entry for APE zip store

[ahgamut]

@jart this PR adds a CosmoImporter as the first entry in sys.meta_path, thereby allowing a glance (and room for optimizations) into another part of Python's import process at the C level.

Entries in sys.meta_path return a ModuleSpec instance if they can find a particular module, otherwise they return None.
With CosmoImporter it is now possible to do things like:

• have a static std::map<modulename, location_in_zip_store> to avoid stat syscalls for modules that are part of the startup process
• prevent unnecessary syscalls/Python operations because we know stdlib modules have to be in the APE zip store

With CosmoImporter, the SourcelessFileLoader added in #408 now has an additional attribute to avoid a stat syscall.

For a rough benchmark I again use the import_stdlib.py from #408, which imports all available packages in the APE stdlib.

with MODE=optlinux and current HEAD commit c06ffd4:

real    0m0.307s
user    0m0.282s
sys     0m0.025s

with MODE=optlinux and current HEAD commit merged with this PR:

real    0m0.297s
user    0m0.289s
sys     0m0.009s

Not much, but by writing better versions of is_builtin and find_frozen and a lookup table for zip store modules, it can be reduced further.

[ahgamut]

All tests pass in MODE=, MODE=tiny, MODE=dbg, MODE=opt,MODE=asan, and MODE=optlinux on my Debian 11 system.

[ahgamut]

Ok I implemented a simple lookup table with binary search to check for modules.
(logarithmic, but each comparison is a strcmp with indirection involved).

It doesn't seem to have improved on the import_stdlib.py benchmark that much, but here's something interesting:

python.com built in MODE= Current HEAD commit 67b28b9

o//third_party/python/python.com --strace -c "2+2" | grep fstat | wc -l
69

with merging this PR it becomes

o//third_party/python/python.com --strace -c "2+2" | grep fstat | wc -l
17

The table ZipCdir_Lookup that I added for zip store modules contains entries for those modules used in APE startup, so we avoid stat calls during startup completely.

[ahgamut]

All python.com tests still pass in MODE=, MODE=tiny, MODE=dbg, MODE=opt,MODE=asan, and MODE=optlinux on my Debian 11 system.

[ahgamut]

CosmoImporter speeds up APE startup by 5% in the default case, as measured via pyperformance:

Benchmark	APE-optlinux-beforeimpmod	APE-optlinux-afterimpmod
python_startup_no_site	9.54 ms	9.06 ms: 1.05x faster
python_startup	14.5 ms	13.9 ms: 1.04x faster
Geometric mean	(ref)	1.05x faster

Another interesting thing I noticed is that APE Python startup time is now competitive with Anaconda Python 3.7:

Benchmark	pyconda37-11	APE-optlinux-afterimpmod
python_startup_no_site	9.79 ms	9.06 ms: 1.08x faster
python_startup	14.0 ms	13.9 ms: 1.00x faster
Geometric mean	(ref)	1.04x faster

Would be nice to get a separate computer to measure benchmarks without jitter.

[ahgamut]

So the above comparisons might now be inaccurate, because I removed the hardcoded table that allowed skipping some stat syscalls :)
But otherwise the changes done here should be ok.