fix(memory): return custom_metadata from VertexAiMemoryBankService.search_memory

[Yatsuiii]

Fixes #6946

Summary

search_memory constructed every returned MemoryEntry with only author, content, and timestamp, silently dropping custom_metadata even though this service's own write path already treats it as first-class: _add_memories_via_create merges MemoryEntry.custom_metadata via _merge_custom_metadata_for_memory before writing, and add_memory's own docstring documents metadata as meaningful, round-tripped data. The write side reads it; the read side never returned it.

Adds _from_vertex_metadata / _from_vertex_metadata_value as the symmetric read-side counterpart to the existing write-side _to_vertex_metadata_value, handling the same four value kinds (bool/double/string/timestamp) rather than only strings.

Why this matters

A caller that writes custom_metadata at creation time -- an application-specific identifier used to correlate a memory back to its source, for example -- had no way to get it back through search_memory. No error, no warning; the field was silently empty on every returned MemoryEntry.

Testing plan

• Verified with a live create -> retrieve round trip against a real Vertex AI Memory Bank instance before implementing the fix, confirming the underlying SDK genuinely returns populated metadata that this class was never reading (details in VertexAiMemoryBankService.search_memory never returns MemoryEntry.custom_metadata, even though the write path treats it as first-class #6946).
• Added test_search_memory_returns_custom_metadata (covers bool/double/string values) and test_search_memory_with_no_metadata_returns_empty_dict (backward-compat: no metadata -> {}, never None), using the existing mock_vertexai_client / _AsyncListIterator fixtures already in the test file.
• Added an assertion to the existing test_search_memory confirming custom_metadata == {} for the pre-existing no-metadata case.
• pytest tests/unittests/memory/ : 89 passed.
• pytest tests/unittests (full suite): 13493 passed, 0 failed (90 skipped / 31 xfailed / 2 xpassed are pre-existing and unrelated).
• pre-commit run --all-files on the changed files: all hooks pass, including the repo's own "ADK Compliance Checks".

Scope note

An earlier draft of #6946 also claimed MemoryEntry.id / memory_id forwarding was broken. Testing that specific claim live contradicted it (forwarding is conditional on an SDK-version compatibility check this file's own code already handles correctly via _get_create_memory_config_keys, and a direct low-level SDK call bypassing that check isn't a fair test of it). That claim was cut from the issue before filing and is out of scope for this PR.