Send Packet Too Big instead of dropping oversized VIP traffic #700

Summary

When the next hop needs fragmentation, the load balancer datapath counts the packet as a drop and discards it. A client whose packet is too big once the 40-byte outer header is added never gets an ICMPv6 Packet Too Big, so its path MTU discovery never adapts and large packets to a VIP vanish.

The NAT shard already generates Packet Too Big (#669), and #641 adds the same for the CNI datapath. Reuse that work here.

Acceptance criteria

• The load balancer sends a rate-limited ICMPv6 Packet Too Big to the client, with an MTU that leaves room for the encapsulation
• The drop is still counted, under its own reason, so the counter shows how often it happens
• A datapath unit test covers it, and the Known Constraints entry in the gateway architecture doc is updated

Phase 0 of #698.