Add interface for register custom ops in torch.ops.ppe

[emcastillo]

Used to avoid decomposition of ops in torch.fx.Graph when using aot_autograd backend, or add new operations with autograd support for new execution backends/devices

# FWD graph traced 
def forward(self, primals_1, primals_2):
    custom_op_fwd = torch.ops.my_custom_library.custom_op_fwd.default(primals_1, primals_1, primals_1, primals_2)
    return [custom_op_fwd, primals_1, primals_2]

# BWD graph traced (name is forward by some pytorch convention)
def forward(self, primals_1, primals_2, tangents_1):
    custom_op_bwd = torch.ops.my_custom_library.custom_op_bwd.default(primals_1, primals_1, primals_1, primals_2);  primals_1 = primals_2 = None
    getitem = custom_op_bwd[0];  custom_op_bwd = None
    return [getitem, None]

[xuzijian629]

I found two small typos. Other code looks good to me!

[xuzijian629]

/test

[xuzijian629]

LGTM!