cuda : enable i16 and i32 for DUP

[amankarki151]

Overview

i32 and i16 DUP ops were silently falling back to the CPU. The i32 path already existed under the hood but was blocked by the support gate. This PR adds the missing i16 branch and fixes the gate to allow both.

Additional information

Testing: test-backend-ops -o DUP passes 10/10 on dual T4s. The full suite passes with zero regressions (16097/16097).

test-backend-ops -o DUP

ggml_cuda_init: found 2 CUDA devices (Total VRAM: 29823 MiB):
  Device 0: Tesla T4, compute capability 7.5, VMM: no, VRAM: 14911 MiB
  Device 1: Tesla T4, compute capability 7.5, VMM: no, VRAM: 14911 MiB
Testing 3 devices

Backend 1/3: CUDA0
  Device description: Tesla T4
  Device memory: 14911 MB (14806 MB free)

  DUP(type=f32,ne=[10,10,20,1]): OK
  DUP(type=f16,ne=[10,10,20,1]): OK
  DUP(type=i32,ne=[10,10,20,1]): OK
  DUP(type=i16,ne=[10,10,20,1]): OK
  DUP(type=f32,ne=[10,10,5,1],permute=[0,2,1,3]): OK
  DUP(type=f16,ne=[10,10,5,1],permute=[0,2,1,3]): OK
  DUP(type=f32,ne=[10,10,5,1],permute=[1,0,2,3]): OK
  DUP(type=f16,ne=[10,10,5,1],permute=[1,0,2,3]): OK
  DUP(type=i16,ne=[10,8,3,1],permute=[0,2,1,3]): OK
  DUP(type=i16,ne=[10,8,3,1],permute=[1,2,0,3]): OK
  10/10 tests passed
  Backend CUDA0: OK
Backend 2/3: CUDA1
  Device description: Tesla T4
  Device memory: 14911 MB (14806 MB free)

  DUP(type=f32,ne=[10,10,20,1]): OK
  DUP(type=f16,ne=[10,10,20,1]): OK
  DUP(type=i32,ne=[10,10,20,1]): OK
  DUP(type=i16,ne=[10,10,20,1]): OK
  DUP(type=f32,ne=[10,10,5,1],permute=[0,2,1,3]): OK
  DUP(type=f16,ne=[10,10,5,1],permute=[0,2,1,3]): OK
  DUP(type=f32,ne=[10,10,5,1],permute=[1,0,2,3]): OK
  DUP(type=f16,ne=[10,10,5,1],permute=[1,0,2,3]): OK
  DUP(type=i16,ne=[10,8,3,1],permute=[0,2,1,3]): OK
  DUP(type=i16,ne=[10,8,3,1],permute=[1,2,0,3]): OK
  10/10 tests passed
  Backend CUDA1: OK
Backend 3/3: CPU
  Skipping CPU backend
3/3 backends passed
OK

Docs: Manually updated the 4 DUP rows in docs/ops/CUDA.csv. I skipped a full regen since running it on a T4 (compute 7.5) would incorrectly downgrade rows for ops that require newer hardware.

Ref: #14909

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - used AI to help navigate the ggml-cuda backend.