Add spyx.experimental.onnx: ONNX export of spiking step functions

to_onnx(model, input_shape, *, batch, dtype, opset, sequence_length) -> bytes
exports a spiking model's single-timestep (x, state) -> (out, new_state) forward to
an ONNX ModelProto via jax2tf -> tf2onnx. ONNX reaches phones (ONNX Runtime Mobile,
NNAPI/CoreML), servers, and edge from one export. tensorflow/tf2onnx imported lazily
(module imports without them; 'pip install tensorflow tf2onnx onnx onnxruntime' to
convert); NOT a declared extra, so the lock/Python-3.14 stay clean.

- Verified: the onnxruntime output matches the JAX single step (out + every new-state
  tensor) at atol/rtol 1e-4 (zero and nonzero state).
- Honest limitation: full-sequence export (sequence_length=T, an ONNX Scan/Loop of the
  whole time loop) does NOT convert here — jax2tf emits a StableHLO XlaCallModule that
  tf2onnx's TF-graph importer can't parse. Documented, test skipped (not faked). Per-
  step export + an external loop is the shipped capability (same shape as LiteRT).

Experimental / unstable API. Third export leg alongside NIR (neuromorphic) and LiteRT.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: kmheckel

src/spyx/experimental/__init__.py

@@ -20,6 +20,11 @@
 - :mod:`spyx.experimental.compress` — bit-packed activation storage for
   memory-efficient BPTT (``packed_spike_dense``, ``pack_spikes``,
   ``unpack_spikes``).
+- :mod:`spyx.experimental.onnx` — export a spiking model to ONNX: the
+  per-timestep step ``(x_t, state) -> (out, new_state)`` (flat tensor I/O), or,
+  when a sequence length is given, the whole ``spyx.nn.run`` temporal loop as a
+  native ONNX ``Scan``/``Loop``. Conversion deps (tensorflow/tf2onnx/onnx) are
+  imported lazily.
 - :mod:`spyx.experimental.stochastic` — stochastic (Bernoulli-spiking) and
   parallelizable prototypes: ``SPSN``, ``StochasticAssociativeLIF``,
   ``StochasticAssociativeCuBaLIF`` and the ``sigmoid_bernoulli`` activations.
@@ -31,7 +36,7 @@
 # here so the experimental surface is discoverable in one place.
 from ..nn import PSU_LIF
 from ..phasor import ResonateFire
-from . import compress, raven, stochastic
+from . import compress, onnx, raven, stochastic
 from .compress import pack_spikes, packed_spike_dense, unpack_spikes
 from .raven import RavenRSM, SlotRouter, SpikingSlotMemory, make_recall_batch
 from .stochastic import (
@@ -44,6 +49,7 @@
 
 __all__ = [
     "compress",
+    "onnx",
     "raven",
     "stochastic",
     "PSU_LIF",