Fix overpermissive percent-decoding of URIs ('%eth1' case)

Invalid `%nn` 3-character strings were being incorrectly decoded. For
example, in `%eth1`, the `%e` was being translated to ASCII 14, while
the correct behavior is to leave `%eth1` unchanged since the string
`%et` is an invalid 3-character percent encoding.

src/core/lib/uri/uri_parser.cc

@@ -46,13 +46,15 @@ std::string PercentDecode(absl::string_view str) {
   std::string unescaped;
   out.reserve(str.size());
   for (size_t i = 0; i < str.length(); i++) {
+    unescaped = "";
     if (str[i] != '%') {
       out += str[i];
       continue;
     }
     if (i + 3 >= str.length() ||
         !absl::CUnescape(absl::StrCat("\\x", str.substr(i + 1, 2)),
-                         &unescaped)) {
+                         &unescaped) ||
+        unescaped.length() > 1) {
       out += str[i];
     } else {
       out += unescaped[0];

test/core/uri/uri_parser_test.cc

@@ -146,6 +146,8 @@ int main(int argc, char** argv) {
                 {{"bar", ""}}, "lol?/");
   test_succeeds("ipv6:[2001:db8::1%252]:12345", "ipv6", "",
                 "[2001:db8::1%2]:12345", {}, {}, "");
+  test_succeeds("ipv6:[fe80::90%eth1.sky1]:6010", "ipv6", "",
+                "[fe80::90%eth1.sky1]:6010", {}, {}, "");
   test_succeeds("https://www.google.com/?a=1%26b%3D2&c=3", "https",
                 "www.google.com", "/", {{"c", "3"}, {"a", "1&b=2"}},
                 {{"a", "1&b=2"}, {"c", "3"}}, "");