Minor improvement to zerotrie hash function

[sffc]

I reviewed our little perfect hash function with @Amanieu who highlighted the modulo operator as likely the performance bottleneck. We brainstormed some various approaches including whether we could clamp the hash table size to a power of 2 so that we could use a bitmask instead of a division. However, there were a few downsides to that approach:

1. This would make the function non-minimal, meaning it would take more storage (mainly directly, but also indirectly by pushing indices to be bigger varints).
2. The non-minimal hash function may require either extra bits to store whether entries exist, or an unfortunate workaround such as storing present keys as the values in the wrong buckets to prevent those entries from being found.
3. A division operator on small integers might not be too expensive depending on the processor.
4. It wasn't clear that this hash function was the bottleneck. Loads of bytes in the zerotrie are what I believe to be the main bottleneck. I've already focused on improving data locality.

What I did in this PR is change the usize division to u8 division. It results in slightly smaller assembly:

Old:
	cmp	rcx, 1
	adc	rcx, 0
	movzx	eax, dl
	cmp	rax, rcx
	jb	.LBB8_9
	xor	edx, edx
	div	ecx

New:

	movzx	eax, cl
	div	dl
	movzx	edi, ah

@Victoronz also pointed me to https://orlp.net/blog/worlds-smallest-hash-table/ which is a good read about perfect hash functions.