Channels last: add the common layout pipeline - #21963

Add ToContiguousChannelsLastPass as the shared fixed-point pipeline for
replacing selected operators with channels-last variants and composing existing
data-movement transforms.

Backends can restrict anchors and provide a node-level propagation barrier.
Reporting covers converted anchors and structural-copy counts and bytes; strict
mode rejects skipped anchors, internal copies, and unknown surviving copy sizes.

Un-skip the proof-of-concept model matrix and drive it through the real
pipeline, keeping both the contiguous and the channels-last case sets, and fold
the focused pipeline coverage into the same suite. The channels-last cases pin
that the pipeline leaves an already-channels-last graph alone.

This optimization loop is transitional. Arm's PropagateViewCopyPermuteDown/Up
is a measurably better mover -- on this same matrix it reaches 82 permutes where
this loop stalls at 156, strictly better in 9 cases and worse in none, with the
wins concentrated in LSTM/GRU/attention. The intent is to generalize that pass
out of backends/arm into backends/transforms and retire _optimization_passes()
into it.

It is not adopted yet because it cannot currently be imported outside Arm:
tosa.RESCALE/TABLE/SCATTER resolve lazily and raise, is_swappable raises rather
than returning False on keepdim != True, and Down crosses an explicit
memory_format on a pointwise-tagged clone without remapping it. It also costs
8.08s against 0.27s here on a 522-node MobileNetV2, and its fan-out invariant
has a five-op counterexample. Remove this loop once those are closed.

Keeping it in the meantime is not cosmetic: on quantized MobileNetV2 through the
explicit-layout path it is the difference between 1 cortex_m::transpose and 106,
with identical convolution counts, so the other 105 are pure NHWC copies at
224x224.

The matrix ends at 156 permutes against 142 before the pass.

Local Corstone-300 run of the full Cortex-M suite, the transforms suites, and
lintrunner.

Authored with Codex.