Add op.input_func to OpInfo-based testing

[IvanYashchuk]

Some functions with a mathematically constrained type of input require special preprocessing so that finite-difference derivatives match the analytical ones implemented in the backward operation. Some examples are Hermitian (or symmetric) linear algebra functions torch.linalg.cholesky, torch.linalg.eigh, or function taking a triangular matrix input as torch.cholesky_inverse.
This requirement can be met for example with a custom OpInfo class that overrides get_op method. But then test_variant_consistency_jit tests would fail with mismatch error in computed gradients because the jit scripted function ignores the get_op method.
OpInfo already has output_func, so we should add input_func as well.

It should be enough to change these lines https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/common_jit.py#L127-L131
to something like

def runAndSaveRNG(self, func, inputs, kwargs=None, input_func=None):
    kwargs = kwargs if kwargs else {}
    with freeze_rng_state():
        if input_func is not None:
            inputs = [input_func(*inputs, **kwargs)]
        results = func(*inputs, **kwargs)
    return results

to make test_variant_consistency_jit tests pass with custom op.input_func. All other occurrences of calling .get_op() should also use op.input_func.

cc @ezyang @albanD @zou3519 @gqchen @pearu @nikitaved @soulitzer @mruberry

[mruberry]

Linked to the test suite tracking issue and cc @albanD as an fyi. This seems like the right idea to me.

[albanD]

SGTM!