Use pinned host memory for GPU state-to-NumPy conversion (#2797)

[mitchdz]

Summary

• Fix performance bottleneck in np.array(cudaq.get_state(kernel)) for GPU-backed
  states by using pinned (cudaMallocHost) host memory instead of pageable (new[])
  for the device-to-host transfer in the pybind11 buffer protocol handler
• Add SimulationState::toHostBuffer() virtual method so GPU backends can provide
  optimized host allocation, with a pageable fallback for non-GPU backends and for
  systems that cannot pin the requested amount of memory
• Fix to_cupy() passing a hardcoded 1024-byte size to UnownedMemory instead of
  the actual buffer size

Motivation

Closes #2797. When np.array() is called on a GPU-resident state, the buffer
protocol handler allocates pageable host memory and copies via cudaMemcpy. With
pageable destinations, CUDA must stage through an internal pinned buffer in chunks,
severely limiting effective bandwidth. On an L4 GPU with 25 qubits (537 MB state):

Allocation	Bandwidth	Time
Pageable (before)	1.8 GB/s	0.29s
Pinned (after)	13.2 GB/s	0.04s

For the original reporter's 32-qubit case on GH200, the ~5.4s overhead should drop
to under 1s.

[mitchdz]

Re-opening as accidentally closed last commit (my bad).

Also, I am not merging this in on purpose in order to do some more proper testing for larger qubit counts. My concern is this new alloc/dealloc may actually cause more overhead.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[mitchdz]

I am blocking this for now. Testing locally on an L4 I do see a regression at higher qubit counts. I did a comparison between main and the new pinned memory logic.

  ┌────────┬────────────┬───────────┬─────────────┬─────────┐
  │ Qubits │ State size │ main (ms) │ pinned (ms) │ Speedup │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     20 │     16 MiB │      8.68 │        6.61 │   1.31x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     21 │     32 MiB │     17.51 │       11.24 │   1.56x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     22 │     64 MiB │     34.06 │       52.25 │   0.65x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     23 │    128 MiB │     67.58 │      102.57 │   0.66x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     24 │    256 MiB │    137.60 │      208.39 │   0.66x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     25 │    512 MiB │    282.67 │      407.51 │   0.69x │
  ├────────┼────────────┼───────────┼─────────────┼─────────┤
  │     26 │      1 GiB │    595.95 │      872.12 │   0.68x │
  └────────┴────────────┴───────────┴─────────────┴─────────┘

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.