fix: handle match_mask == 0 in NEON ctzll to avoid undefined behavior

When all 16 bytes match the allowed range, match_mask becomes 0 after
the bitwise NOT. Calling __builtin_ctzll(0) is undefined behavior.

The code expects match_len == 16 when all bytes match (so the branch
is skipped and p += 16 continues the loop), but this relied on
ctzll(0) returning 64, which is not guaranteed.

Example panic on macOS ARM64:

    thread 44856 panic: passing zero to ctz(), which is not a valid argument
    src/llhttp/llhttp.c:2654:21: in llhttp__internal__run
            match_len = __builtin_ctzll(match_mask) >> 2;
                        ^

Fix by explicitly checking for match_mask == 0 and setting match_len = 16.

Author: lalinsky

src/implementation/c/node/table-lookup.ts

@@ -245,7 +245,13 @@ export class TableLookup extends Node<frontend.node.TableLookup> {
     // https://community.arm.com/arm-community-blogs/b/servers-and-cloud-computing-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
     out.push('  narrow = vshrn_n_u16(mask, 4);');
     out.push('  match_mask = ~vget_lane_u64(vreinterpret_u64_u8(narrow), 0);');
-    out.push('  match_len = __builtin_ctzll(match_mask) >> 2;');
+    // When all 16 bytes match, match_mask is 0. Calling __builtin_ctzll(0) is
+    // undefined behavior, so we handle this case explicitly.
+    out.push('  if (match_mask == 0) {');
+    out.push('    match_len = 16;');
+    out.push('  } else {');
+    out.push('    match_len = __builtin_ctzll(match_mask) >> 2;');
+    out.push('  }');
     out.push('  if (match_len != 16) {');
     out.push(`    ${ctx.posArg()} += match_len;`);
     {