perf: optimize PartialEq implementation

[klkvr]

Due to most significant limbs being filled first comparison is suboptimal right now for smaller values

[codspeed-hq]

Merging this PR will improve performance by 29.41%

⚡ 4 improved benchmarks
✅ 102 untouched benchmarks
⏩ 111 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	eq/8	538.9 µs	416.4 µs	+29.41%
⚡	eq/16	538.9 µs	416.4 µs	+29.41%
⚡	eq/32	511.1 µs	416.4 µs	+22.73%
⚡	eq/64	511.1 µs	416.4 µs	+22.73%

---

_{Comparing klkvr/partial-eq-perf (8df38fb) with main (ff87f1c)}

Footnotes

1. 111 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[DaniPopes]

@klkvr can you show asm diff

[klkvr]

@DaniPopes amp got me this

--- a/a.s
+++ b/b.s
@@ -4,13 +4,32 @@
 	.type	nibbles_eq,@function
 nibbles_eq:
 	.cfi_startproc
-	vmovdqu	(%rdi), %ymm0
-	vmovq	32(%rdi), %xmm1
-	vmovq	32(%rsi), %xmm2
-	vpxor	%ymm2, %ymm1, %ymm1
-	vpxor	(%rsi), %ymm0, %ymm0
-	vpor	%ymm1, %ymm0, %ymm0
-	vptest	%ymm0, %ymm0
-	sete	%al
-	vzeroupper
+	mov	rcx, qword ptr [rdi]
+	cmp	rcx, qword ptr [rsi]
+	jne	.LBB0_4
+	mov	rax, qword ptr [rdi + 32]
+	cmp	rax, qword ptr [rsi + 32]
+	jne	.LBB0_4
+	mov	al, 1
+	cmp	rcx, 17
+	jae	.LBB0_6
+.LBB0_3:
 	ret
+.LBB0_6:
+	mov	rdx, qword ptr [rdi + 24]
+	cmp	rdx, qword ptr [rsi + 24]
+	jne	.LBB0_4
+	cmp	rcx, 33
+	jb	.LBB0_3
+	mov	rdx, qword ptr [rdi + 16]
+	cmp	rdx, qword ptr [rsi + 16]
+	jne	.LBB0_4
+	cmp	rcx, 49
+	jb	.LBB0_3
+	mov	rax, qword ptr [rdi + 8]
+	cmp	rax, qword ptr [rsi + 8]
+	sete	al
+	ret
+.LBB0_4:
+	xor	eax, eax
+	ret

[DaniPopes]

currently it's branchless and doesn't matter how long the value is. with this change it's probably slightly faster for when the workload is only small values, even then i doubt it