inflate more SIMD

[folkertdev]

needs refinement, and just hardcodes x86 + AVX2.

The general gist here is that we get fewer instructions, but the same number of cycles as a safer version without this much SIMD. But that can be very instruction-set dependent. So we should have more instruction sets before we can really make this decision.

[folkertdev]

I tried to improve the offset_from_end == 1 case, but it is only faster in the micro benchmark, and actually slower in the inflate of silesia-small.tar.gz

            if offset_from_end == 1 {
                use std::arch::x86_64::{_mm256_set1_epi8, _mm256_storeu_si256};

                // this will just repeat this value many times
                let element = self.buf[current - 1];
                let b = unsafe { element.assume_init() };

                if false && safe_to_chunk {
                    let chunk = unsafe { std::arch::x86_64::_mm256_set1_epi8(b as i8) };
                    for d in self.buf[current..][..length].chunks_mut(32) {
                        unsafe {
                            _mm256_storeu_si256(d.as_mut_ptr().cast(), chunk);
                        }
                    }
                } else {
                    self.buf[current..][..length].fill(element);
                }
            } else {
                for i in 0..length {
                    self.buf[current + i] = self.buf[start + i];
                }
            }