net: dsa: make assisted_learning_on_cpu_port bypass offloaded LAG interfaces

Given the following topology, and focusing only on Box A:

         Box A
         +----------------------------------+
         | Board 1         br0              |
         |             +---------+          |
         |            /           \         |
         |            |           |         |
         |            |         bond0       |
         |            |        +-----+      |
         |192.168.1.1 |       /       \     |
         |  eno0     swp0    swp1    swp2   |
         +---|--------|-------|-------|-----+
             |        |       |       |
             +--------+       |       |
               Cable          |       |
                         Cable|       |Cable
               Cable          |       |
             +--------+       |       |
             |        |       |       |
         +---|--------|-------|-------|-----+
         |  eno0     swp0    swp1    swp2   |
         |192.168.1.2 |       \       /     |
         |            |        +-----+      |
         |            |         bond0       |
         |            |           |         |
         |            \           /         |
         |             +---------+          |
         | Board 2         br0              |
         +----------------------------------+
         Box B

The assisted_learning_on_cpu_port logic will see that swp0 is bridged
with a "foreign interface" (bond0) and will therefore install all
addresses learnt by the software bridge towards bond0 (including the
address of eno0 on Box B) as static addresses towards the CPU port.

But that's not what we want - bond0 is not really a "foreign interface"
but one we can offload including L2 forwarding from/towards it. So we
need to refine our logic for assisted learning such that, whenever we
see an address learnt on a non-DSA interface, we search through the tree
for any port that offloads that non-DSA interface.

Some confusion might arise as to why we search through the whole tree
instead of just the local switch returned by dsa_slave_dev_lower_find.
Or a different angle of the same confusion: why does
dsa_slave_dev_lower_find(br_dev) return a single dp that's under br_dev
instead of the whole list of bridged DSA ports?

To answer the second question, it should be enough to install the static
FDB entry on the CPU port of a single switch in the tree, because
dsa_port_fdb_add uses DSA_NOTIFIER_FDB_ADD which ensures that all other
switches in the tree get notified of that address, and add the entry
themselves using dsa_towards_port().

This should help understand the answer to the first question: the port
returned by dsa_slave_dev_lower_find may not be on the same switch as
the ports that offload the LAG. Nonetheless, if the driver implements
.crosschip_lag_join and .crosschip_bridge_join as mv88e6xxx does, there
still isn't any reason for trapping addresses learnt on the remote LAG
towards the CPU, and we should prevent that.

Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>