strings: fix performance of nextind

The recursion (for invalid bytes) was preventing inlining, as was the
length of the function. For ASCII data, the cost of the call far exceeds
the cost of decoding the data.

Closes #51624

base/strings/string.jl

@@ -177,9 +177,16 @@ end
     @boundscheck between(i, 1, n) || throw(BoundsError(s, i))
     @inbounds l = codeunit(s, i)
     (l < 0x80) | (0xf8 ≤ l) && return i+1
+    (@noinline function _nextind_continued(s, i, n, l) # mark the rest of the function as a slow-path
     if l < 0xc0
+        # handle invalid codeunit index by scanning back to the start of this index
+        # (which may be the same as this index)
         i′ = @inbounds thisind(s, i)
-        return i′ < i ? @inbounds(nextind(s, i′)) : i+1
+        i′ >= i && return i+1
+        i = i′
+        @inbounds l = codeunit(s, i)
+        (l < 0x80) | (0xf8 ≤ l) && return i+1
+        @assert l >= 0xc0
     end
     # first continuation byte
     (i += 1) > n && return i
@@ -192,7 +199,8 @@ end
     ((i += 1) > n) | (l < 0xf0) && return i
     # third continuation byte
     @inbounds b = codeunit(s, i)
-    ifelse(b & 0xc0 ≠ 0x80, i, i+1)
+    return ifelse(b & 0xc0 ≠ 0x80, i, i+1)
+    end)(s, i, n, l)
 end
 
 ## checking UTF-8 & ACSII validity ##