Fix QLinearSoftmax output for empty inputs (#33401)
### Description
<!-- Describe your changes. -->
Create the `QLinearSoftmax` output before returning early for an empty
input, matching the existing CPU `Softmax` behavior. The nonempty
computation is unchanged.
Add CPU regression tests for int8/uint8, opsets 12/13, and zero-sized
dimensions, with and without shape metadata and with constant or runtime
quantization parameters.
Tested on Windows x64 Release:
- The two new tests fail before the fix and pass afterward.
- All 18 lookup-table operator tests pass.
- A broader Softmax run passes 52 tests; 7 CUDA/WebGPU tests are skipped
in this CPU-only build.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Fixes #33262.
For an empty input such as `[0, 2, 37]`, the kernel returned success
without creating an output tensor. A downstream `DequantizeLinear` node
then failed with `Missing Input`. Creating the output first preserves
its shape and type while still skipping the arithmetic when there are no
elements.
Co-authored-by: David Luong <luongdavid@microsoft.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 3f6d4277-d757-4418-9b0e-91c7c1fab8f6
