[CUDA] Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy

[gaugarg-nv]

Add a CUDA ggml_cpy fast path for same-type, same-shape strided copies that are just 2D pitched block copies. When tensors are not fully contiguous, but each row is contiguous, it now uses cudaMemcpy2DAsync instead of the slow element-wise scalar copy kernel.

This fixes the GDN recurrent snapshot update with -np 4, where rollback slots are separated by cache stride gaps. This has no impact on -np 1 perf as that is already using optimized cudaMemcpyAsync path.

MTP=3 on RTX 5090 with Unsloth Qwen3.6-27B Q4_K_M with -np 4 shows a perf gain of ~7.5%.

Master:

python3 mtp-bench.py
  code_python        pred= 192 draft= 162 acc= 137 rate=0.846 tok/s=147.0
  code_cpp           pred= 192 draft= 179 acc= 130 rate=0.726 tok/s=132.6
  explain_concept    pred= 192 draft= 189 acc= 127 rate=0.672 tok/s=126.5
  summarize          pred= 192 draft= 166 acc= 134 rate=0.807 tok/s=141.8
  qa_factual         pred= 192 draft= 178 acc= 131 rate=0.736 tok/s=134.1
  translation        pred= 192 draft= 186 acc= 129 rate=0.694 tok/s=129.5
  creative_short     pred= 192 draft= 201 acc= 122 rate=0.607 tok/s=117.3
  stepwise_math      pred= 192 draft= 179 acc= 131 rate=0.732 tok/s=133.0
  long_code_review   pred= 192 draft= 206 acc= 122 rate=0.592 tok/s=115.7

PR:

python3 mtp-bench.py
  code_python        pred= 192 draft= 162 acc= 137 rate=0.846 tok/s=157.9
  code_cpp           pred= 192 draft= 179 acc= 130 rate=0.726 tok/s=142.4
  explain_concept    pred= 192 draft= 189 acc= 127 rate=0.672 tok/s=135.9
  summarize          pred= 192 draft= 166 acc= 134 rate=0.807 tok/s=151.8
  qa_factual         pred= 192 draft= 178 acc= 131 rate=0.736 tok/s=143.8
  translation        pred= 192 draft= 186 acc= 129 rate=0.694 tok/s=139.2
  creative_short     pred= 192 draft= 201 acc= 122 rate=0.607 tok/s=126.3
  stepwise_math      pred= 192 draft= 179 acc= 131 rate=0.732 tok/s=143.5
  long_code_review   pred= 192 draft= 206 acc= 122 rate=0.592 tok/s=124.7

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, paired with Codex.

[am17an]

Can we add a test-backend-ops for this shape in case it doesn't exist?

[gaugarg-nv]

Can we add a test-backend-ops for this shape in case it doesn't exist?

Thanks. There indeed was no test executing this path. I have added one now.

[gaugarg-nv]

OpenVino backend is failing these two tests. Adding checks in OpenVino to return unsupported for strided copies.
OpenVino doesn't hit this for GDN as the K > 1 (MTP) case is already unsupported.