Speed up Image.getextrema()

[akx]

Speeds up getextrema() -- with more than just hoist and restrict this time, by adding a separate function for the common case of 2-4 band 8bpc images. The single-channel case speeds up pleasantly too by autovectorisation, though.

------------------- benchmark 'extrema': 4 tests, 2 sources --------------------
Name (time in us)                   0001_3a36c3a OPS  0003_796ae94 OPS      ΔOPS
--------------------------------------------------------------------------------
test_getextrema[1024x1024-RGB]            1,790.4542       13,099.6057   +631.6%
test_getextrema[1024x1024-L]              1,410.9957       17,641.2525  +1150.3%
test_getextrema[1024x1024-RGBA]             687.5769       13,102.1316  +1805.6%
test_getextrema[1024x1024-LA]               494.9731       11,851.6516  +2294.4%
--------------------------------------------------------------------------------

[codspeed-hq]

Merging this PR will improve performance by ×3

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 6 improved benchmarks
✅ 597 untouched benchmarks
⏩ 335 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_getextrema[1237x811-CMYK]	20.3 ms	4.7 ms	×4.3
⚡	test_getextrema[1237x811-RGBA]	20.3 ms	4.7 ms	×4.3
⚡	test_getextrema[1237x811-LA]	15.7 ms	4.7 ms	×3.3
⚡	test_getextrema[1237x811-RGB]	14.2 ms	4.7 ms	×3
⚡	test_getextrema[1237x811-L]	3.6 ms	1.4 ms	×2.5
⚡	test_getextrema[1237x811-I]	6.7 ms	4.3 ms	+55.06%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing akx:extremaly-fast (6002bd5) with main (0d20a00)}

Footnotes

1. 335 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[radarhere]

Considering that we don't have any images with more than 4 bands, I'm not sold that we need to check for it.

[akx]

If this check is elided, in the future when someone decides to add such support, the following code will crash in unfortunate and non-obvious ways when it writes over the UINT8 mb[2 * 4]; array, where 4 stands for "maximum number of bands supported by this code".

I can remove the check if you like, but I don't think it costs much here.

[radarhere]

Couldn't you just allocate mb to match the number of bands?

[akx]

Using a VLA (or heap allocation) becomes a little complex, since due to the data layout of 2-band (PA/LA/La) images (pixelsize 4, XX....AA in image32), the actual valid sizes for mb (and ImagingGetExtremaMultiband's vmin and vmax) are actually only 3 or 4. IOW, the required accumulator array size is not directly related to bands.

I'm not sure it's worth the complexity, what do you think?