ENH: Add 1D convolutional flows (ConvFlow1d).

[ntustison]

Normalizing flows for multichannel sequences of shape (B, C, L), e.g. 2D keypoint trajectories over a gait cycle.

New components:

• nets.ConvNet1d: 1D conv subnet (zero-init last layer, optional ActNorm, configurable padding_mode).
• flows.Invertible1x1Conv1d: LU-parameterized invertible 1x1 conv for NCL tensors (log-det scaled by L), reusing the existing guardrails.
• flows.CouplingBlock1d: channel-split affine coupling with bounded log-scales (s_cap * tanh) and optional shift cap.
• flows.GlowBlock1d: ActNorm -> Invertible1x1Conv1d -> coupling, with configurable kernel_size and the same nan/gen_clamp guards as 2D/3D.
• ConvFlow1d: single-scale stack of K coupling or Glow blocks over a DiagGaussian((C, L)) base.
• Permute(mode="reverse").

Design notes:

• padding_mode="circular" makes the couplings equivariant to circular time shifts (periodic signals); default stays "zeros" as in 2D/3D.
• actnorm_s_cap (default 5.0) is decoupled from scale_cap (0.5) so ActNorm can standardize unscaled inputs.
• ConvFlow1d.sample() accepts temperature (as MultiscaleFlow.sample) and restores the base temperature afterwards.

Tests perturb all weights (zero-init couplings are otherwise the identity) and check, in float64, non-trivial transforms, round trips, log-det against the autograd Jacobian, log_prob consistency, circular equivariance, ActNorm standardization and tempered sampling.