Potentially optimizing `build_huffman_tree`

[okaneco]

I came across this article and was wondering if an approach like this was tried for the following tree building code with repeated SIMD min lookups instead of maintaining a min-heap.
https://www.intel.com/content/www/us/en/developer/articles/technical/fast-computation-of-huffman-codes.html

fdeflate/src/compress/bitstream.rs

Lines 68 to 70 in cbaf854

	let mut lengths = [0u8; 286];
	let mut codes = [0u16; 286];
	build_huffman_tree(&frequencies, &mut lengths, &mut codes, 15);

If I understand correctly, they found above ~140 elements seemed to be where the vector method was preferred over the heap. However, they're using AVX2 and u16 instead of u32 which I think the current code uses. I'm not sure if it's possible to efficiently get the min position. with implicit autovectorization.

I don't know the impact this would have on overall runtime since they numbers they report are just for the tree building.

[fintelia]

I haven't tried something like that so it could be worth experimenting with. Though probably worth checking how long we currently spend on tree building before putting too much effort into it.

Edit: Did some profiling and found that building the Huffman tree takes a couple percent of the total encode time for the faster levels and a fraction of a percent for the slower ones.