Switch to lazy exports

[flying-sheep]

On my systems, this speeds up import:

• 1.26s→0.91s on Linux (1.38× speed improvement)
• 1.47s→1.07s on macOS (1.37× speed improvement)

(in an environment with everything except Cupy installed)

TODO:

• figure out lazy singledispatch

[codecov]

Codecov Report

Attention: Patch coverage is 95.69892% with 4 lines in your changes missing coverage. Please review.

Project coverage is 96.86%. Comparing base (a7e216b) to head (a83784b).

Files with missing lines	Patch %	Lines
src/fast_array_utils/_import.py	90.69%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #22      +/-   ##
==========================================
- Coverage   96.95%   96.86%   -0.10%     
==========================================
  Files          11       12       +1     
  Lines         164      223      +59     
==========================================
+ Hits          159      216      +57     
- Misses          5        7       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[codspeed-hq]

CodSpeed Performance Report

Merging #22 will improve performances by ×3.6

_{Comparing pa/lazy (a83784b) with main (a7e216b)}

Summary

⚡ 2 improvements
✅ 50 untouched benchmarks

Benchmarks breakdown

	Benchmark	BASE	HEAD	Change
⚡	test_stats_benchmark[1-scipy.sparse.csr_array-float32-is_constant]	417.3 µs	115 µs	×3.6
⚡	test_stats_benchmark[1-scipy.sparse.csr_array-float64-is_constant]	250.1 µs	148.5 µs	+68.43%

[ilan-gold]

So the goal of this is to speed up imports? Looking at the code, I'm not sure it's worth this level of complexity TBH.

Things like:

@_sum.register("dask.array:Array")

make me cringe a bit. I would need to let this sit with me for a bit before I'd give it a sign-off. I'm very hesitant

[flying-sheep]

The syntax is actually a Python standard used in entry points, e.g. scripts and all others that I know of, e.g. pygments lexers, pytest plugins, web services (WSGI/ASGI, e.g. uvicorn main:app), …

[ilan-gold]

I meant more creating a custom register based on this, not the syntax itself of the string.

[ilan-gold]

I think honestly that I don't see the tradeoff here worth it IMO. I think the complexity of this sort of thing can make our code unapproachable to people who might want to contribute. I understand if the complexity is really necessary here, but I don't personally feel it is needed here :/