Use ObjectEmpty() instead of array-only Empty() on a JSON object

[Daksha1611]

🛠 Summary

Fixes #4547.

computeDeltaImpl() builds its result with delta.SetObject(), so nestedDelta is always an object, but it was tested with Value::Empty(). In the pinned RapidJSON that accessor is array-only:

// rapidjson/document.h:1649
bool Empty() const { RAPIDJSON_ASSERT(IsArray()); return data_.a.size == 0; }
// rapidjson/document.h:1197
bool ObjectEmpty() const { RAPIDJSON_ASSERT(IsObject()); return data_.o.size == 0; }

RAPIDJSON_ASSERT is assert() and is not overridden anywhere in the tree, so a build with assertions live (-c dbg, Makefile:118) aborts with Assertion `IsArray()' failed as soon as a streamed delta contains a nested object present in both snapshots — ordinary streaming of a tool call with nested arguments. Under NDEBUG the assert disappears and data_.a.size aliases data_.o.size, so the wrong accessor returns the right number. Release behaviour is unchanged by this PR.

The four tool parsers computing the same kind of delta already use ObjectEmpty() (hermes3, llama3, mistral, phi4), so this brings the builder in line. The other two Value::Empty() call sites in src/ are genuinely arrays and are left alone.

The five delta assertions in partial_json_builder_test.cpp made the same misuse, so the suite aborted under -c dbg before reaching the production call; they now use ObjectEmpty(). Adds computeDeltaWithUnchangedNestedMember for the branch whose emptiness result decides the output. Note that test passes both before and after in a release build, for the aliasing reason above — its value is running correctly under assertions.

I have no OVMS build container available, so this is not compiled against the full tree and CI will need to confirm the build. I did reproduce the abort and the fix against the exact pinned RapidJSON revision from WORKSPACE:102. Draft for that reason.

🧪 Checklist

• Unit tests added.
• The documentation updated.
• Change follows security best practices.