refactor(io): split FDA TxDevice/HVController from RUO subclasses

[openlifu-agent]

Closes #43
Closes #44
Refs OpenwaterHealth/openlifu-diathermy-application#16 (SR-002)

Stacked on #45 (base feature/SR-002-device-interface). Merge #45 first.

What

DeviceInterface (the FDA entry point) handed out TxDevice and
HVController objects that still had every RUO command on them.
Restricting the interface alone left those commands one attribute away.
This PR moves the split down to the component level.

FDA (DeviceInterface)	RUO (LIFUInterface)
TxDevice: identity, temperatures, get_trigger/start_trigger/stop_trigger, get_preset/load_preset, metadata	LIFUTxDevice(LIFUComponentMixin, TxDevice): set_trigger*, profiles/set_solution, registers, async_mode, module invert, firmware update
HVController: identity, HV on/off/status, get_preset/select_preset, telemetry reads	LIFUHVController(LIFUComponentMixin, HVController): set_voltage, DACs, 12 V, fan, RGB, hv_enable
OWComponent: connect/close, send_checked, ping/version/HWID/read_config	LIFUComponentMixin: uart, start/stop, send_async, echo, LED, reset, DFU, write_config*

• Opcode whitelist: each FDA class has SUPPORTED_COMMANDS
  (TX_FDA_COMMANDS / HV_FDA_COMMANDS in LIFUConfig).
  send/send_checked raise ValueError for any other opcode
  before the UART is touched, so a raw send_checked(OW_TX7332_WREG)
  on an FDA object is refused.
• Async is RUO-only: DeviceInterface no longer has run_async,
  start_monitoring/stop_monitoring, or async_mode calls in
  start_sonication/stop_sonication. LIFUInterface keeps all
  of these.
• Hooks: _create_devices (LIFUInterface builds the LIFU
  classes) and _connect_devices (LIFUInterface skips the sync
  connect in async mode).
• Simulator: mirrors the split (SimulatedLIFUTxDevice,
  SimulatedLIFUHVController; monitoring and async_mode are on
  SimulatedLIFUInterface).

Behaviour changes to note

• LIFUInterface.start_sonication now sets async_mode before
  turning HV on (previously after). Both happen before the trigger
  starts, so the sonication sequence is unchanged.
• The FDA DeviceInterface has no monitor thread, so there is no
  USB hot-plug auto-reconnect on the FDA path. Ports are opened once
  in __init__.
• TxDevice.__init__ no longer takes module_invert (it is now on
  LIFUTxDevice).

Residual gaps (not fixed here)

• OW_CMD_USR_CFG covers both read and write (reserved=1). The
  whitelist checks only the opcode, so it cannot block a hand-built
  config write. write_config itself is no longer on the FDA classes.
• _uart is private by convention only.
• The whitelist ignores packet type. This is safe today because command
  values are unique across packet types.
• The operator interface's firmware-update flow has no real-hardware
  path on DeviceInterface. The sim keeps an update_firmware
  placeholder. Tracked in Add a service interface for firmware update / config writes outside the FDA DeviceInterface #46 (service interface).

Test plan (executed)

• SDK: pytest unit-test gives 223 passed, 113 subtests, with no deselections. The baseline before this change was 147 passed with test_tx_device.py ignored entirely.
• New tests in test_fda_presets.py:
  • TestComponentSurface: the FDA classes lack every RUO member, the
    LIFU classes have them, and the simulator classes match.
  • TestCommandWhitelist: FDA TX rejects OW_TX7332_WREG and
    OW_CMD_ECHO, FDA HV rejects OW_POWER_SET_HV, and none of
    these calls reaches send_packet. The LIFU classes allow them.
  • TestInterfaceSurface: DeviceInterface._create_devices builds
    exactly TxDevice/HVController, and
    LIFUInterface._create_devices builds the LIFU classes. Monitoring
    is now in RUO_ENDPOINTS.
• Operator interface (feature/SR-002-device-interface + this SDK):
  all 12 test files pass. test_sonication_progress.py needed a
  test-only change because it used set_trigger on the FDA sim; that
  change is on OI PR #124 (commit c4762c3).

#44

test_tx_device.py exercises the class being split, so this PR fixes
its HW_ID_DATA_LENGTH import (now from LIFUConfig) and points it
at LIFUTxDevice. Once the file collected, three tests
(test_23b/23c/23d) failed. They asserted the 2.0.14
train-interval auto-fill and the bump for firmware 2.0.3 and older.
set_trigger has sent 0 for back-to-back trains since 8b3eaf2,
and firmware 2.0.3 and older no longer needs support. A separate commit
replaces those three tests with tests of the current contract (23b-23f):

• Zero or zero-gap intervals are sent as 0.
• Longer intervals pass through unchanged.
• A short interval is sent as 0 for a single train, and raises
  ValueError with nothing sent for multiple trains.

Self-review

• Assumption I'm making: the FDA app needs only the opcodes in
  TX_FDA_COMMANDS/HV_FDA_COMMANDS. That set is derived from
  every method the operator interface calls on txdevice/hvcontroller
  (verified by its test suite).
• Alternative considered: putting the RUO classes in separate
  modules. I kept them next to their FDA bases to keep the diff
  reviewable and preserve import paths. The FDA classes still do not
  define any RUO method.
• New dependency introduced: none
• Test coverage: SR-002 component surface, opcode whitelist, and
  interface factories. Not covered: real-hardware async/monitor path
  (unchanged code, moved), and the FDA no-hot-plug behaviour.
• Breaking change: code that called RUO methods on a
  DeviceInterface's components, or passed run_async /
  async_mode to DeviceInterface, must use LIFUInterface. The
  only in-tree caller of that kind was an OI test (fixed).
  LIFUInterface users are unaffected.
• Tech-debt filed: Add a service interface for firmware update / config writes outside the FDA DeviceInterface #46 (service interface). Closes test_tx_device.py fails to import HW_ID_DATA_LENGTH #44.