Bypass BufReader and state machine when decoding

[fintelia]

This sketches out how we could switch away from our current BufRead approach of taking slices of input to drive a state machine. It instead takes a Read impl and pull bytes from it as needed.

Leaving this draft state because there's APNG decoding isn't fully working at the moment and because I'm not fully sure we want to go down this path of duplicating the decoding logic between the low-level streaming API and the higher-level decoding API.

Performance is mixed with both significant improvements and regressions. It seems to vary based on the constants driving the minimum and maximum IDAT buffer sizes passed to the zlib decompressor.

decode/kodim17.png
                        time:   [-2.7008% -2.6371% -2.5847%] (p = 0.00 < 0.05)

decode/kodim07.png
                        time:   [+0.6333% +1.3371% +2.5938%] (p = 0.00 < 0.05)

decode/kodim02.png
                        time:   [+0.8983% +1.0315% +1.1734%] (p = 0.00 < 0.05)

decode/kodim23.png
                        time:   [+0.1002% +0.1620% +0.2178%] (p = 0.00 < 0.05)

decode/Lohengrin_-_Illustrated_Sporting_and_Dramatic_News.png
                        time:   [-1.0207% -0.6634% -0.3293%] (p = 0.00 < 0.05)

decode/Transparency.png time:
                        time:   [+77.678% +78.337% +78.794%] (p = 0.00 < 0.05)

generated-noncompressed-4k-idat/8x8.png
                        time:   [-44.694% -44.589% -44.510%] (p = 0.00 < 0.05)

generated-noncompressed-4k-idat/128x128.png
                        time:   [+28.799% +29.139% +29.482%] (p = 0.00 < 0.05)
...

generated-noncompressed-2g-idat/12288x12288.png
                        time:   [-21.915% -20.856% -19.853%] (p = 0.00 < 0.05)

[anforowicz]

Is it fair to say that the buffer inside BufReader is replaced with the idat_stream: Vec<u8>?

[fintelia]

Yeah, I think that's a reasonable summary. And then #458 extends this further to eliminate the idat_stream Vec