Partially revert "CXX-2120 handle nonexistent document fields (#984)"

[eramongodb]

A small precursor change to CXX-3234 which partially reverts #984 in advance of the upcoming v_noabi refactor, which will adopt v1's "last known element key" behavior instead.

---

Note

The CXX-2120 implementation is private API whose resulting (undocumented) behavior is the format of error messages, which we do not (want to) consider part of stable API compatibility requirements (although this intention is not explicitly documented by our API Versioning policy... it should really be documented in a yet-to-be-written "Compatibility Guidelines" document). Therefore, this change in behavior is not considered to be an API breaking change.

As noted by comments in prior PRs (here and here), v1 implements an alternative approach to CXX-2120 which is not prone to dangling issues caused by the internal string view. Although (hopefully) unlikely, this dangling issue can be demonstrated by the following code:

// Current API behavior:
bsoncxx::document::value doc = bsoncxx::from_json(R"({"x": 1})");

// Intended behavior: improved exception messages.
try {
  auto type = doc["y"].type();
} catch (bsoncxx::exception const& ex) {
  CHECK_THAT( // Lookup key is present. :)
    ex.what(),
    Catch::Matchers::ContainsSubstring(R"("y")")
  );
}

// Problem: behavior is not extended to `element` access.
try {
  auto type = doc["x"]["y"].type();
} catch (bsoncxx::exception const& ex) {
  CHECK_THAT( // Neither 'x' nor 'y' are present. :(
    ex.what(),
    !Catch::Matchers::ContainsSubstring(R"("x")") && !Catch::Matchers::ContainsSubstring(R"("y")")
  );
}

// Problem: element object may be long-lived (by design).
{
  auto y = doc[std::string("y")]; // Undocumented internal long-term storage as stdx::string_view.
  //           ^~~~~~~~~~~~~~~~   // Temporary object!

  // AddressSanitizer: stack-use-after-scope
  //     # 0 in strlen
  //     # ...
  //     # 2 in std::string::string(char const*)
  //     # 3 in bsoncxx::v_noabi::document::element::type() const
  //     # ...
  CHECK_THROWS(y.type());
}

Due to this potential dangling issue (+ the optional<T>'s impact on element's otherwise semitrivialty), this PR reverts these changes in advance of upcoming changes as part of CXX-3234 which will adopt v1's "last known element key" behavior instead:

// API behavior following CXX-3234:
bsoncxx::document::value doc = bsoncxx::from_json(R"({"x": 1})");

try {
  auto type = doc["y"].type();
} catch (bsoncxx::exception const& ex) {
  CHECK_THAT( // Unfortunately no lookup key... :(
    ex.what(),
    !Catch::Matchers::ContainsSubstring(R"("y")")
  );
}

try {
  auto type = doc["x"]["y"].type();
} catch (bsoncxx::exception const& ex) {
  CHECK_THAT( // ... but does have "last known element key" instead. :)
    ex.what(),
    Catch::Matchers::ContainsSubstring(R"("x")")
  );
}

{
  auto y = doc[std::string("y")]; // No internal reference semantics to the key argument.
  CHECK_THROWS(y.type()); // OK: no undefined behavior.
}

The five ContainsSubstring() assertions modified by this PR will be modified once more by the CXX-3234 PR. These will be the only assertion modifications made by CXX-3234: there will be no other observable change in v_noabi behavior.