layer: ResNet-style skip connections for 1D and 2D CNNs

[LeoBuron]

Feature

ResNet-style residual blocks for 1D and 2D CNNs: identity shortcut around conv→norm→act→conv→norm with an elementwise-add merge, plus the projection shortcut (kernel-size-1 conv) for blocks that change channel count or stride.

Blockers

• training_loop: non-sequential model topology — branch/merge DAG for forward+backward (epic) #329 (topology epic) — branch/merge forward+backward is the hard prerequisite: Add-merge node, gradient accumulation at the fork.
• 2D variant additionally blocked by layer/arithmetic: 2D sliding-window layer family for image recognition (epic) #330 (2D layer family). The 1D variant has no other missing kernels.

Notes

• The 1D variant should ship first: the Conv1d family and Add arithmetic exist, so a residual block is the minimal consumer of training_loop: non-sequential model topology — branch/merge DAG for forward+backward (epic) #329's Add-merge — a good acceptance vehicle for the epic itself.
• Projection shortcut = Conv1d/Conv2d with kernel size 1 — no new kernel needed.
• SYM configs: the merge-add of two differently-scaled SYM wires depends on training_loop: non-sequential model topology — branch/merge DAG for forward+backward (epic) #329's quantized-merge semantics; keep the first residual example FLOAT32 until that spec lands.

Acceptance

A small residual CNN example (HAR or ECG variant) with a PyTorch twin; trajectory comparison over ≥10 seeds per repo rule.