selfdrive/car: ban cereal and capnp

[sshane]

instead of #33197

while thinking about it, I realized we might not need to switch to a full blown class in the first iteration. will try to reduce touching things, and just convert log.CanData to a custom dataclass with generic send/receive functions passed into fingerprinting/car interface code

[sshane]

We have builder, automatic defaults, and proper static analysis type checking now! Operates similarly to Enum's auto() object.

This plus the custom enum helper is only +37 lines for capnp struct behavior but with static linting, I think that's worth it.

@auto_dataclass
class TestClass:
  a: int = auto_field()
  b: float = auto_field()
  c: str = auto_field()
  d: bytes = auto_field()
  e: list[int] = auto_field()
  f: tuple[int] = auto_field()
  g: set[int] = auto_field()
  h: dict[str, int] = auto_field()
  i: bool = auto_field()
  ecu: CarParams.Ecu = auto_field()
  carFw: CarParams.CarFw = auto_field()
  carFwList: list[CarParams.CarFw] = auto_field()

TestClass()
Out[3]: TestClass(a=0, b=0.0, c='', d=b'', e=[], f=(), g=set(), h={}, i=False, ecu=<Ecu.eps: 'eps'>, carFw=CarParams.CarFw(ecu=<Ecu.unknown: 'unknown'>, fwVersion=b'', address=0, subAddress=0, responseAddress=0, request=[], brand='', bus=0, logging=False, obdMultiplexing=False), carFwList=[])

[sshane]

process replay diffs:

A lot are minute differences due to switching from float32 precision to float64 using Python's float type. Nissan calculates mean of floats to get steering pressed for example. Toyota rate limit being different was because its CarControllerParams didn't call which as a function, and the <torque which-enum> defines a __str__/__eq__

I made a Which StrEnum that supports __call__ which is more robust:

CP.lateralTuning.which == 'torque'
Out[6]: True
CP.lateralTuning.which() == 'torque'
Out[7]: True

Previously, which would return a method.

---

Here's a sample from the Subaru for example:

	('change', 'carState.steeringRateDeg', (-1.0849952697753906, -1.0850000381469727))
	('change', 'carState.aEgo', (0.37458106875419617, 0.37458622455596924))

Setting NUMPY_TOLERANCE to 7e-6 2e-5 resolves all diffs (Subaru's angle rate calculation is the highest diff).

[sshane]

@adeebshihadeh I've reviewed the car interface code, can you review card.py and structs.py? For the car tests touched here, I am going to leave them in openpilot for now, but I can see splitting them up like I did for the docs test in a follow up PR (not going to do here since they're all so huge and require some more thought).

[sshane]

We unfortunately have to rewrite these to take advantage of referencing vs. deepcopying and avoiding two additional type()/isinstance() on every single field recursively, which take around ~7-8% total vs 3-4% with this