fix(server): surface a typed error when world_engine is requested on Apple Silicon

[juraam]

Partially addresses #81 — reports the incompatibility at session init rather than ahead of time.

On Apple Silicon devices.py forces WORLD_ENGINE_DEVICE="cpu", so the world_engine backend
loads the model and the seed successfully and only dies at the second warmup step
(append_frame), where _cache_pass is compiled:

NotImplementedError: torch.compile on current platform is not supported for CPU. target: flex_attention

The user sits through a full model load and warmup to get a compiler internals dump, with no
hint that the backend setting is the thing to change.

Before

After

I checked whether the CPU path is worth unblocking rather than rejecting: with
TORCHDYNAMO_DISABLE=1 warmup does complete on an M1 Pro — append_frame 1658 s, gen_frame
3140 s, ~80 min total, ~0.0013 pixel FPS. Functionally correct, nowhere near playable.

Fix: raise a typed IncompatibleHardwareError before the model load and surface it as
MessageId.INCOMPATIBLE_HARDWARE over both the error push and the deferred init rpc_err.

The pre-init socket is deliberately not torn down — it acks the init and keeps waiting, like
the not ready branch beside it. Closing in the same tick made the client's lifecycle reducer
see a plain connection failure before the RPC rejection landed, so the precise message lost the
race and only appeared on a second attempt.

Tested on an M1 Pro dev build. Lints, basedpyright and codegen_ts.py --check pass locally.