feat: properly model acquisition time constraints

[vijk777]

We add two modes:

• time_aligned: data available at 0, tu, 2tu, ...
• staggered_random: each neuron has data at an interval of tu like above. But the neurons have random acquisition offsets 0, 1, ..., tu-1. This is like the zapbench data.

  time_aligned mode (tu=5)
  ============================
  All neurons observed simultaneously at regular intervals

  Time:     0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20
           ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
  Neuron 1 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│
  Neuron 2 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│
  Neuron 3 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│
  Neuron 4 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│
  Neuron 5 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│
           └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
           ↑              ↑              ↑              ↑              ↑
         t=0            t=5           t=10           t=15           t=20


  staggered_random mode (tu=5, different phases per neuron)
  ==========================================================
  Each neuron observed at different phase + k*tu

  Time:     0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20
           ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
  Neuron 1 │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  phase=0
  Neuron 2 │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │  phase=1
  Neuron 3 │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  phase=3
  Neuron 4 │  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  phase=4
  Neuron 5 │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  │ X│  │  │  │  phase=2
           └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
           ↑  ↑  ↑  ↑  ↑                                                
        phases 0-4 spread across first tu timesteps
        (then repeat every tu steps)

Deprecate intermediate loss points.

Intermediate results are quite promising for time-aligned acquisition (middle), while naively applying the current model to staggered acquisition fails (right). The left column is the baseline where we use all the data. And I've selected epoch 20/100. We see errors of ~ 1e-2 and I am also showing the constant baseline and we are 100x better than that.

This is different from what we saw on the old DAVIS data - where using time-aligned acquisition resulted in poor generalization and a divergence.