fix: use rotated mask register in scalar byte tails

The scalar tail (.m_bytes, .u_bytes, .gf_bytes) used [rsi+1] and
[rsi+2] (original mask pointer) for the last 2-3 bytes instead of
the rotated r8d register. After AVX2 alignment prologues rotate the
mask for non-multiple-of-4 offsets, these produce wrong XOR bytes.

Fix all three tails to advance r8d via ror and use r8b consistently.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: mvogttech

src/ws_mask_asm.asm

@@ -500,13 +500,15 @@ ws_mask:
     mov [rdx], al
     dec rcx
     jz .m_ret
+    ror r8d, 8
     mov al, [rdi + 1]
-    xor al, byte [rsi + 1]
+    xor al, r8b
     mov [rdx + 1], al
     dec rcx
     jz .m_ret
+    ror r8d, 8
     mov al, [rdi + 2]
-    xor al, byte [rsi + 2]
+    xor al, r8b
     mov [rdx + 2], al
 .m_ret:
     ret
@@ -928,12 +930,12 @@ ws_unmask:
     xor byte [rdi], r8b
     dec rcx
     jz .u_ret
-    mov al, byte [rsi + 1]
-    xor byte [rdi + 1], al
+    ror r8d, 8
+    xor byte [rdi + 1], r8b
     dec rcx
     jz .u_ret
-    mov al, byte [rsi + 2]
-    xor byte [rdi + 2], al
+    ror r8d, 8
+    xor byte [rdi + 2], r8b
 .u_ret:
     ret
 
@@ -2098,13 +2100,15 @@ ws_mask_gfni:
     mov [rdx], al
     dec rcx
     jz .gf_ret
+    ror r8d, 8
     mov al, [rdi + 1]
-    xor al, byte [rsi + 1]
+    xor al, r8b
     mov [rdx + 1], al
     dec rcx
     jz .gf_ret
+    ror r8d, 8
     mov al, [rdi + 2]
-    xor al, byte [rsi + 2]
+    xor al, r8b
     mov [rdx + 2], al
 .gf_ret:
     ret