serviceability: add new_interfaces vec to Device with custom serializer

[elitegreg]

Part of the forward-compatible Device interfaces refactor.
Design: https://gist.github.com/elitegreg/e1d27c97034656a980fa8d7628ae16c5

Depends on #3657 (NewInterface struct).

Scope

Append new_interfaces: Vec<NewInterface> to Device after max_multicast_publishers, behind a custom BorshSerialize that keeps the existing interfaces: Vec<Interface> (legacy enum vec) length-locked with the new vec via per-write projection.

The legacy field stays at its current byte position so old readers don't change.

Tasks

state/device.rs

• Drop #[derive(BorshSerialize)] on Device.
• Append pub new_interfaces: Vec<NewInterface> as the last field.
• Custom BorshSerialize for Device:
  • Rebuild legacy interfaces from new_interfaces via InterfaceV2::from(&NewInterface) wrapped as Interface::V2(...) before serializing (always V2 to match the post-smartcontract: revert default Interface to V2; keep V3 for migrate/backfill #3653 on-disk current).
  • Serialize all fields in declaration order. The legacy vec lands at its existing offset; new_interfaces lands at the end.
  • assert_eq!(legacy.len(), new_interfaces.len()).
• Update TryFrom<&[u8]> for Device (lines 481-523):
  • Read legacy interfaces at the existing offset.
  • Read all post-interface fields up through max_multicast_publishers.
  • Try to read trailing Vec<NewInterface> with unwrap_or_default() so legacy accounts (no trailing bytes) yield Vec::new().
  • If trailing vec is empty, populate new_interfaces from legacy via the per-variant TryFrom<&InterfaceV1/V2> for NewInterface conversions. Otherwise use the trailing vec as-is.
  • If len(trailing) != len(legacy), panic — should not happen outside the empty-trailing case.
• Update Default impl to initialize both vecs as Vec::new().
• Device::find_interface returns (usize, &NewInterface) (was (usize, CurrentInterfaceVersion)). Note: ~30 call sites ripple — those move in subsequent issues, not this one. Where blockers arise, add a temporary helper that returns the legacy projection so this PR stays self-contained.

Tests

• test_device_serialize_keeps_vecs_in_sync: round-trip a Device with N entries; assert len(legacy) == len(new_interfaces) == N and per-element name equality.
• test_device_legacy_account_rebuilds_new_vec: hand-serialized device omitting the trailing vec — Device::try_from populates new_interfaces from legacy.
• test_device_skips_future_interface_in_new_vec: forge version=5 on the first new_interfaces element with N trailing junk bytes; both interfaces remain accessible.
• Existing test_state_compatibility_device continues to pass and now exercises the legacy fallback.

Notes

• No processor changes in this issue. Keeping both vecs in sync is the responsibility of the device serializer, so write-path callers in subsequent issues can mutate new_interfaces and rely on the projection.