[QNN-EP] add gather handler for transpose optimizer

[quic-muchhsu]

Description

Gather is missing from the transpose optimizer's handler map. This PR adds HandleGather, which pushes a Transpose past a Gather so the
transpose can cancel or fuse with neighboring transposes.

The rewrite is:

data -> Transpose(perm) -> Gather(axis=k, indices=scalar_const)

becomes

data -> Gather(axis=perm[k], indices=scalar_const) -> Transpose(SqueezePerm({perm[k]}, perm))

Scope is intentionally narrow — only the scalar (0-D) constant-indices case, which is structurally identical to a Squeeze along the gathered axis and reuses SqueezePerm for the post-rewrite output perm. Non-scalar or non-constant indices are left untouched. The framework's DefaultCostCheck still gates the rewrite on profitability, so unprofitable cases are rejected before the handler runs.

Tests added in transpose_optimizer_test.cc cover the positive scalar-indices case, negative-axis normalization, rank-1-indices fall-through, and dynamic-indices fall-through.

Motivation and Context

We've observed that the Transpose → Gather → Transpose pattern produces sub-optimal inference performance on the QNN EP. With this
handler in place, the surrounding transposes can fold or cancel, eliminating layout-conversion overhead around Gather.

[Copilot]

Pull request overview

Adds ONNX Gather support to the transpose optimizer so a leading Transpose can be pushed past a Gather when indices is a 0-D constant scalar, enabling subsequent transpose fusion/cancellation (notably improving QNN EP graphs with Transpose→Gather→Transpose patterns).

Changes:

• Implement HandleGather in the transpose optimizer to remap axis under the incoming perm, push the transpose through, and emit the correct reduced-rank output permutation via SqueezePerm.
• Register the new handler in the optimizer’s ONNX handler map.
• Add unit tests covering: positive scalar-indices rewrite, negative-axis normalization, and two explicit no-opt fall-through cases (rank-1 indices and non-constant indices).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/optimizer/transpose_optimization/onnx_transpose_optimization.cc	Adds a narrow Gather handler for scalar constant indices and registers it in the default handler map.
onnxruntime/test/optimizer/transpose_optimizer_test.cc	Adds coverage for the new Gather rewrite and key non-applicable cases.

[edgchen1]

hm, the windows_x64_asan build keeps failing consistently:

https://github.com/microsoft/onnxruntime/actions/runs/28128962920/job/83556848419?pr=28755#step:5:20897

1: AddressSanitizer: Out of memory. The process has exhausted 8192MB for size class 8192.
1: =================================================================
1: ==11460==ERROR: AddressSanitizer: allocator is out of memory trying to allocate 0x2000 bytes
1: AddressSanitizer: nested bug in the same thread, aborting.

I'm not sure why. maybe try merging from main first? if it still is an issue, there might be something to investigate further.