[Operator Mechanism] refine erfinv out-of-domain NaN follow-up (#78820)

- Use typed quiet_NaN<float16|bfloat16>() in GPU functor specializations
  instead of casting from float NaN, leveraging Paddle's numeric_limits
  specializations.
- Wrap dygraph block in test_erfinv_op.py with try/finally so
  paddle.enable_static() always runs, preventing global mode leakage on
  assertion failure.
- Extend GPU dygraph test to cover float16 (and bfloat16 when supported),
  exercising the float16/bfloat16 functor specializations in CI.