This fix PR #22817 introduced performance regression.

In that pull request, the input length check was moved forward,
but the related ori instruction was missing, and it will cause
input of any length down to the much slower scalar implementation.

Fixes #23300

CLA: trivial

crypto/chacha/asm/chacha-loongarch64.pl

@@ -71,6 +71,7 @@
 	# $a4 = arg #5 (counter array)
 
 	beqz		$len,.Lno_data
+	ori			$t3,$zero,64
 	la.pcrel	$t0,OPENSSL_loongarch_hwcap_P
 	ld.w		$t0,$t0,0
 
@@ -461,7 +462,6 @@ sub ROUND_4x {
 $code .= <<EOF;
 .align 6
 .LChaCha20_4x:
-	ori			$t3,$zero,64
 	addi.d		$sp,$sp,-128
 
 	# Save the initial block counter in $t4
@@ -886,7 +886,6 @@ sub ROUND_8x {
 $code .= <<EOF;
 .align 6
 .LChaCha20_8x:
-	ori			$t3,$zero,64
 	addi.d		$sp,$sp,-128
 
 	# Save the initial block counter in $t4