perf: use swar to skip whitespace

[cyruspyre]

I don't know if you will accept it but, using swar to skip whitespace improves performance when skipping over large amount of whitespaces. But also degrades, when there is none at all. I tried to avoid the degradation but was unable to do it. I've tried checking for 1-2 chars outside the loop but the generated asm was doing that already and manually doing it somehow made it even worse.

                      old             new        change
-- with -C target-cpu=native --------------------------
twitter/struct    753.11 MiB/s    804.69 MiB/s   +6.85%
twitter/dom       245.85 MiB/s    248.78 MiB/s   +1.19%
citm/struct       787.24 MiB/s    958.22 MiB/s  +21.72%
citm/dom          384.28 MiB/s    427.41 MiB/s  +11.22%
canada/struct     588.49 MiB/s    558.73 MiB/s   -5.06%
canada/dom        148.92 MiB/s    136.03 MiB/s   -8.65%

-- without --------------------------------------------
twitter/struct    749.26 MiB/s    796.89 MiB/s   +6.36%
twitter/dom       248.85 MiB/s    249.94 MiB/s   +0.44%
citm/struct       844.54 MiB/s    992.05 MiB/s  +17.46%
citm/dom          394.14 MiB/s    427.33 MiB/s   +8.42%
canada/struct     581.38 MiB/s    546.50 MiB/s   -6.00%
canada/dom        149.54 MiB/s    139.65 MiB/s   -6.61%

Running with -C target-cpu=native just made it use YMM register to perform the same operations...

[cyruspyre]

I was somewhat able to reduce the slow down when there is no whitespace by replacing the code in end_seq function with loop equivalent and using simple loop to parse whitespace instead.

There was swar code duplication and a call to panic_in_cleanup in the asm of end_seq. I don't know why there was a call to such function at all. My best guess is that it has something to do with the existence of ErrorCode::Message(Box<str>). Nonetheless, doing so improved/degraded perf again.

                      old             new        change
-- with -C target-cpu=native --------------------------
twitter/struct    753.11 MiB/s    825.14 MiB/s   +9.56%
twitter/dom       245.85 MiB/s    247.15 MiB/s   +0.53%
citm/struct       787.24 MiB/s    960.71 MiB/s  +22.04%
citm/dom          384.28 MiB/s    416.97 MiB/s   +8.51%
canada/struct     588.49 MiB/s    566.53 MiB/s   -3.73%
canada/dom        148.92 MiB/s    139.45 MiB/s   -6.36%
-- without --------------------------------------------
twitter/struct    749.26 MiB/s    823.83 MiB/s   +9.95%
twitter/dom       248.85 MiB/s    251.76 MiB/s   +1.17%
citm/struct       844.54 MiB/s    998.68 MiB/s  +18.25%
citm/dom          394.14 MiB/s    434.81 MiB/s  +10.32%
canada/struct     581.38 MiB/s    576.26 MiB/s   -0.88%
canada/dom        149.54 MiB/s    140.62 MiB/s   -5.97%