perf: channels-last Conv2d dwconv for LynxNet2/ConvNeXt, AdamW routing for dwconv - #341

…g for dwconv

- LynxNet2 blocks and the ConvNeXt decoder keep the [B, T, C] layout end to
  end: the depthwise conv runs as channels-last Conv2d over zero-copy views
  of the Conv1d weights, avoiding the transposes plus the hidden contiguity
  copies Conv1d makes on transposed input (full-model fp16 fwd+bwd measured
  10-23% faster for T <= 4096 frames)
- eager falls back to the original Conv1d path above 4096 frames (cuDNN
  depthwise backward regresses on long sequences); graph export always emits
  the Conv2d form (ORT CPU EP runs 1D depthwise conv 2-17x slower than the
  2D form, DirectML is indifferent)
- new AdamWDWConv1d marker subclass routes depthwise conv weights to AdamW
  instead of Muon (same mechanism as AdamWLinear/AdamWConv1d), applied to
  LynxNet, LynxNet2 and ConvNeXt
- state_dict compatible with existing checkpoints; Triton fused-kernel patch
  for LynxNet2 (softsign GLU) stays compatible; exported graphs verified
  equivalent to the previous Conv1d graphs on ORT CPU/DML

…bit nitpick)

The trace -> jit.script -> ONNX export flow degraded reshape(int-list) and
squeeze into dynamic Shape/Gather/Concat chains and guarded onnx::If nodes
that onnxsim cannot fold inside the sampling Loop subgraph. Replace them
with single static ops: unsqueeze for the 4D input/weight views and
select(2, 0) to drop the always-1 height dim. A dwconv segment is now just
Unsqueeze/Unsqueeze/Conv/Gather in the exported graph.

Verified with the full official acoustic export (conditioner extraction,
onnxsim x3, graph merge): ORT CPU loads and runs the merged model, outputs
identical to the pre-change Conv1d graphs (CPU & DML, T=781/3125/10000).

Per review feedback the ONNX export must stay on the original 1D Conv
graphs, so the channels-last path now runs only while training with
T <= 4096 frames; evaluation, inference and graph export always use the
plain Conv1d computation. Verified the exported graphs are identical to
the pre-change graphs (op sequence and all initializers) with the official
acoustic and variance export pipelines, and that ORT CPU/DML outputs are
bit-identical.

The weight is back to the plain 3D (C, 1, K) shape: state_dict keys and
shapes are unchanged, checkpoints stay compatible in both directions, and
no state_dict hooks are involved.

Training speedups are unchanged (1.07-1.20x on the production configs).

The fast path expanded the weight with unsqueeze(1), which only produces
the required (C_out, C_in / groups, 1, K) shape when C_in / groups == 1
(depthwise). For grouped or dense convs it put the input-channel dim into
the height position and failed with a shape error. Use unsqueeze(2) so the
kernel always spans the width (T) dim, and cover the grouped/dense cases in
the tests.

Also reject string padding and padding_mode != 'zeros' in the constructor:
the fast path maps the Conv1d kernel onto the Conv2d width dim and cannot
honour either of them.