[QDP] Add zero-copy amplitude encoding from float32 GPU tensors

[viiccwen]

Purpose of PR

This PR adds encode_from_gpu_ptr_f32 and encode_from_gpu_ptr_f32_with_stream to QdpEngine, enabling zero-copy amplitude encoding from float32 GPU pointers. It relies on the existing GpuStateVector Float32 support and the launch_amplitude_encode_f32 / launch_l2_norm_f32 kernels.

Summary

• New APIs (Linux/CUDA):
  • QdpEngine::encode_from_gpu_ptr_f32(input_d, input_len, num_qubits) — amplitude encoding from GPU f32 pointer using the default stream.
  • QdpEngine::encode_from_gpu_ptr_f32_with_stream(..., stream) — same with an explicit CUDA stream (null = default).
• GPU pointer validation: All encode_from_gpu_ptr and encode_batch_from_gpu_ptr paths now validate the input pointer via cudaPointerGetAttributes (non-null, device/managed memory, same device as engine). Early checks for empty input / sample size are added where missing.
• Amplitude encoder: calculate_inv_norm_gpu_f32 is refactored to call a new calculate_inv_norm_gpu_f32_with_stream; the stream-aware variant is used for f32 GPU encoding and is synchronized before host copy.
• Python binding: get_torch_cuda_stream_ptr returns null when PyTorch reports stream pointer 0 (default stream) instead of raising, so default-stream usage is supported.
• Tests: New Rust tests for encode_from_gpu_ptr_f32 and encode_from_gpu_ptr_f32_with_stream (default and non-default stream, f32/f64 engine, empty input, input length > state size). Existing GPU pointer tests updated to use valid GPU pointers so they still hit the intended error paths after the new validation. Basis tests use Float64 engine. Test helper create_test_data_f32 added in tests/common.
• Python tests: Redundant in-function import torch removed; test_encode_cuda_tensor_unsupported_encoding parametrized with iqp / iqp-z and error message match updated to the current supported list.

## BEFORE: f64 path (old API — user had to convert f32→f64 to use this)
## AFTER: f32 path (new API — zero-copy from PyTorch f32 CUDA tensors)

Encode-from-GPU-pointer benchmark: 16 qubits, 200 iterations, state_len=65536
  BEFORE (encode_from_gpu_ptr f64): 0.0838 ms/encode, 11933.2 encodes/s
  AFTER  (encode_from_gpu_ptr_f32): 0.0632 ms/encode, 15820.1 encodes/s
  Speedup (f32 vs f64 path): 1.33x

Related Issues or PRs

closes #996 , also a follow-up PR for #995

Changes Made

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Breaking Changes

• Yes
• No

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes
• Successfully built and ran all unit tests or manual tests locally
• PR title follows "MAHOUT-XXX: Brief Description" format (if related to an issue)
• Code follows ASF guidelines

[guan404ming]

Design looks nice to me. Could you help attach before and after benchmark on your local machine to prove your improvement works? Thanks!

[viiccwen]

@guan404ming got it!

## BEFORE: f64 path (old API — user had to convert f32→f64 to use this)
## AFTER: f32 path (new API — zero-copy from PyTorch f32 CUDA tensors)

Encode-from-GPU-pointer benchmark: 16 qubits, 200 iterations, state_len=65536
  BEFORE (encode_from_gpu_ptr f64): 0.0838 ms/encode, 11933.2 encodes/s
  AFTER  (encode_from_gpu_ptr_f32): 0.0632 ms/encode, 15820.1 encodes/s
  Speedup (f32 vs f64 path): 1.33x

[guan404ming]

Looks nice, please help resolve the conflicts. Thanks!

[viiccwen]

cc @guan404ming @rich7420
updated PR description

[guan404ming]

Looks nice for the change.
Sorry about that but conflict comes again.

[rich7420]

@viiccwen thanks for the update!
I think we could add a test like e.g. test_encode_from_gpu_ptr_f32_null_pointer or somthing that calls encode_from_gpu_ptr_f32(std::ptr::null(), len, qubits) and asserts Err(MahoutError::InvalidInput(msg)) with msg.contains("null").
then overall lg

[rich7420]

LGTM

[viiccwen]

Thx all!