Stream explicit GRU state

// Explicit state creates a detached graph boundary for streaming and TBPTT.
import aner.nn
import aner.tensor

let x = tensor.reshape(nn.random(2, 4, seed: 7, scale: 0.2), List(2, 4, 1))
let y = nn.zeros(2, 1)
let model = nn.gru(1, 8).on("cpu")
let start = model.initial_state(batch: 2)
let first = model.step(x, start, lengths: List(4, 2))
let next = model.step(x, first.state())
print(next.values().rows())
let fitted = model.fit_chunk(x, y, start, rate: 0.001, clip_norm: 1.0)
let continued = fitted.fit_chunk(x, y, fitted.state())
print(continued.updates())
// Keep stream identity/order stable across chunks. Reset at series boundaries.