Transforms: let a backend declare what counts as a permute

[rascani]

The propagation pass matched a single hardcoded target, aten.permute_copy, so a
backend carrying its own layout dialect got a pass that silently did nothing: it
never recognised the copies, so it never moved them.

Take the set through the constructor, and thread it into the three passes the
driver delegates its actual work to. Recognising a permute widens to the
declared set; creating one still emits _PERMUTE_TARGET, so a caller that passes
no argument sees the behaviour it had before.

Threading it all the way down is the part that matters. The driver walks the
graph itself but cancels nothing on its own -- fuse_vertical hands off to
CanonicalizeViewCopyPermutePass and fuse_horizontal to
FuseIdenticalInputTransformsPass, both of which hardcoded aten.permute_copy.
Parameterising only the driver leaves it walking correctly over copies that the
passes underneath cannot see. Measured on Cortex-M explicit-layout mv2, counting
cortex_m::transpose:

driver alone, driver-only parameterisation      106   (unoptimized baseline)
driver alone, threaded through                   31
driver + region cancellation, threaded through    1
the Cortex-M fixed-point loop it would replace    1

So this is what makes the shared driver reach parity on a second backend, and
make_fusion_pass is the composition point that closes the remaining 31.

Found by wiring the driver into the Cortex-M pipeline and measuring rather than
by reading: every other extension point is exercised by the Arm suite, and this
one is not, because Arm has only one permute target.

No behaviour change for Arm: the pass suite produces a failure set identical to
merge base, name for name.

Authored with Claude Code.

[rascani]

Stack from ghstack (oldest at bottom):

• -> Transforms: let a backend declare what counts as a permute #22125
• Transforms: let the shared driver decline a reduction it cannot cross #22124
• Transforms: share the permute/view propagation driver #22123
• Transforms: let the shared driver decline a reduction it cannot cross #22226

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22125

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 280 Pending

As of commit 33053b2 with merge base 35adf60 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: rascani / name: RJ Ascani (251721d, 33053b2, 627e259, 67880ae, 710f0f3, 7b7eb72, 8a3739e, 8b2dd09, a54e6e7, a726ffd, be06ff4, d46fe9c, dda250f, fd0c239)