Add SynapticDelay: per-channel axonal / synaptic delay with a learnable delay time

Closes #60.

Motivation

snnTorch has no delay mechanism. A spike emitted at step t always reaches
the next layer at step t. There is no way to give a pathway a
transmission lag, and no way to learn one — even though axonal/synaptic
delays are a standard feature of biological circuits and, since Hammouamri,
Khalfaoui-Hassani & Masquelier (Learning Delays in Spiking Neural Networks
using Dilated Convolutions with Learnable Spacings, ICLR 2024), a
competitive trainable parameter for temporal tasks.

What this adds

One module, snn.SynapticDelay — a plain nn.Module placed between a
connection and a neuron:

self.fc1   = nn.Linear(num_in, num_hidden)
self.delay = snn.SynapticDelay(max_delay=16, delay=1.0,
                               channels=num_hidden, learn_delay=True)
self.lif1  = snn.Leaky(beta=0.9)
...
cur1 = self.delay(self.fc1(x))          # step mode: (batch, N)
spk1, mem1 = self.lif1(cur1, mem1)

• Fractional, trainable delay. Each channel is delayed by its own
  d_c ≥ 0, realised by interpolation so d_c is continuous.
  kernel="linear" — exact 2-tap kernel, local gradient (±1 step), best
  for fine delays / fine-tuning. kernel="gaussian" — gradient across a
  wide delay range; anneal sigma from wide to narrow during training to
  learn a large delay from scratch (the DCLS recipe).
• Two call modes, one kernel. Step mode (batch, channels), called per
  time step, keeps a ring buffer and a reset_delay() that mirrors
  Leaky.reset_mem(). Sequence mode (time, batch, channels) runs the
  whole sequence at once as a causal depthwise conv1d along time — pairs
  with LeakyParallel. The two modes give the same output for the same
  delay.
• snnTorch conventions. learn_delay toggles nn.Parameter vs
  register_buffer; the delay is clamp(0, max_delay)-ed every forward,
  as beta is clamped to [0, 1].
• Additive only. No existing neuron signature changes. SynapticDelay
  does not subclass SpikingNeuron (it has no membrane / threshold), so it
  is reset explicitly via reset_delay() rather than through
  utils.reset() — same pattern as calling Leaky.reset_mem() yourself.

Evidence

Pre-registered acceptance bars (gradcheck, RED/GREEN, honest baselines):

17 new unit tests in tests/test_snntorch/test_synapticdelay.py.

Scope / follow-ups (deliberately out)

• Per-synapse (weight-matrix-shaped) delays, the full DCLS form — this is
  per-channel. A [out, in] kernel is a natural follow-up.
• Delay in the recurrent path of RLeaky / RSynaptic.
• NIR export of the delay parameter.