Add a benchmark for forward+backward for defuns.

Also fix some simple issues that I saw when I benchmarked it (goes from ~3500 examples/sec -> ~4000 examples/sec)

- (nest) Expose is_mapping check that caches to python.
- (nest) Stop calling flatten when unnecessary in pack_sequence_as
- (nest) Set some functions to their swig wrappers directly (instead of wrapping them in another function)
- Directly call the gen_math_ops call in _aggregate_grads to skip any unnecessary python overhead.
- Stop falling back to slow path in _fast_fill.

PiperOrigin-RevId: 209223633