fix(a2a): read long-running function name from data, not metadata

[Osamaali313]

Fixes #6295

Problem

LongRunningFunctions._mark_long_running_function_call sets the A2A task state for a long-running function call — auth_required for an End-User-Credential (EUC) request, input_required otherwise. It read the function name from the wrong dict:

# src/google/adk/a2a/converters/long_running_functions.py:163
if a2a_part.root.metadata.get("name") == REQUEST_EUC_FUNCTION_CALL_NAME:

For a function-call DataPart, the name lives in a2a_part.root.data (the FunctionCall dump — see part_converter.convert_genai_part_to_a2a_part: data=part.function_call.model_dump(...)). metadata only holds the ADK adk_type / adk_is_long_running keys, so metadata.get("name") is always None. The EUC branch is never taken, and a credential request is mislabeled input_required instead of auth_required.

The sibling converter already does this correctly:

# src/google/adk/a2a/converters/event_converter.py:504
and part.root.data.get("name") == REQUEST_EUC_FUNCTION_CALL_NAME

Fix

if a2a_part.root.data.get("name") == REQUEST_EUC_FUNCTION_CALL_NAME:

Tests

Added tests/unittests/a2a/converters/test_long_running_functions.py (there was no dedicated test — the executor test fully mocks LongRunningFunctions), asserting that an EUC request yields auth_required and a regular long-running function yields input_required.

Before the fix the EUC case fails (input_required); after, it passes. The full tests/unittests/a2a/converters/ suite passes (171 → 173 with the new tests).