Carry the weight-layout tag through an Identity fold

[enthropy7]

Stacked on #25 — targets onnx-optimizer-ir, not main.

The tracker model does not load on #25: Grouped Conv channel mismatch: IC=3, OC=3, group=16 at /xif1_0/dw/conv_1/Conv. Its weight holds a pre-permuted
depthwise [3, 3, 16, 1] that no layout table lists, so the Conv parses it as
[O, I, KH, KW] and reads the kernel height back as the output-channel count.
main loads and runs the same model. yolo11n was unaffected, which is why it
did not show up sooner.

Two halves of one round trip were missing:

• foldable_inputs refused any node with a pre-permuted operand. Right for an
  operator that interprets the bytes, wrong for Identity, which only copies
  them — refusing it lets the node survive to run time and hand its consumer a
  permuted tensor with no tag. This model reads the weight twice, once straight
  (which is what registers it) and once through an alias, so only the second
  reader broke.
• apply_ir deliberately left the layout side tables alone, on the grounds that
  passes preserve weight names. A folded alias is a new name. They are now
  rebuilt from the IR's per-value weight_layouts, which survives the rename.

Also gates the three plan tests that assume the pre-permuted layout. They pin a
fused plan shape that only forms when the loader permutes, and
cargo test -p yscv-onnx --features gpu --lib (ci.yml:86) keeps ONNX-native
OIHW, where the plan legitimately differs — a guaranteed red on that job rather
than a flake. Reproduced locally at 243 passed / 3 failed, matching CI exactly.

The new regression test fails without the fold change and passes with it.

Tracker back to 8.86 ms/1T on Zen 4 against 8.79 on main, so the IR rework does
not cost anything on this model once it loads. Workspace suite 2373 passed;
clippy clean with default features and with gpu.

[enthropy7]

Closing — landing directly on main instead.