[QDP] feat: add credit card fraud benchmark + amplitude encoding optimizations

[rich7420]

Changes

• New benchmark: encoding_benchmarks/qdp_pipeline/creditcardfraud_amplitude.py — 5-qubit
  amplitude VQC on Credit Card Fraud data, aligned with PennyLane baseline (same circuit, loss,
  optimizer). Closes the QDP vs baseline training time gap from ~22% slower to <1% gap.
• New baseline: encoding_benchmarks/pennylane_baseline/creditcardfraud_amplitude.py —
  PennyLane reference implementation with AUPRC/F1 metrics for imbalanced data.
• QuantumDataLoader API: added source_array(X) (in-memory, no temp file),
  as_torch(device), and as_numpy() for ergonomic batch output format.
• Rust PipelineIterator: added new_from_array() constructor; InMemory next_batch now
  passes &data[start..end] slice directly (no per-batch to_vec()).
• amplitude.rs: moved D2H norm validation to after encode kernel + device.synchronize(),
  eliminating a mid-pipeline GPU→CPU roundtrip in encode_batch.
• Bug fixes (iris + creditcard benchmarks): requires_grad=False on all data arrays to
  prevent AdamOptimizer from computing unnecessary gradients through state vectors;
  AmplitudeEmbedding(normalize=False) in place of StatePrep; .real extraction after
  torch.from_dlpack() to convert complex128 DLPack output to float64.

Motivation

The existing QDP benchmark suite only covers the Iris dataset (100 samples, 2 qubits), which is too small to surface real-world data-loading and encoding bottlenecks. Credit Card Fraud (284,807 transactions, 5 qubits) is a standard imbalanced-classification benchmark from Kaggle/OpenML that stresses the full QDP pipeline — batch iteration, GPU encoding, and training — at realistic scale.

Adding this benchmark serves this purposes:

• Expands loader API coverage: source_array(X), as_torch(), and as_numpy() are exercised end-to-end by the new benchmark and tests, catching integration issues across the Python → PyO3 → Rust → CUDA boundary.

Checklist

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other
• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[rich7420]

It seems a little be too big one. Sorry about that.

[viiccwen]

Thx for contributing! 🙌
I'll look deeper tomorrow, and I think we should add tests to cover new loader APIs, especially the new behavior crosses Python, PyO3, Rust, and CUDA boundaries.

[viiccwen]

as_torch() validates that torch is installed, but as_numpy() does not. Then _wrap_iterator() calls _torch.from_dlpack(...) for the "numpy" path.

Does it mean as_numpy() can succeed at configuration time and then fail during iteration with an unclear runtime error if PyTorch is not installed. 🤔

[rich7420]

oh , nice catch! you're right

[ryankert01]

amplitude.rs: moved D2H norm validation to after encode kernel + device.synchronize(),
eliminating a mid-pipeline GPU→CPU roundtrip in encode_batch.

nice

[400Ping]

Please solve conflicts

[rich7420]

plz take a look and test, not hurry once you have time

[ryankert01]

The pennylane baseline uses CPU for training, whereas qdp pipeline has dual CPU/GPU. We can simplify both to use GPU for training always.

[rich7420]

no problem!

[400Ping]

Please solve conflicts

[400Ping]

Can you also give some more context in the PR description for why is this needed?

[400Ping]

The array-loader optimization claim does not match the implementation in this PR.
create_array_loader() says batching uses slices without per-batch to_vec(), but
PipelineIterator::take_batch_from_source() still clones each in-memory batch with
data[start..end].to_vec().

[rich7420]

The array-loader optimization claim does not match the implementation in this PR. create_array_loader() says batching uses slices without per-batch to_vec(), but PipelineIterator::take_batch_from_source() still clones each in-memory batch with data[start..end].to_vec().

Great catch! You are right that the code in take_batch_from_source() was misleading.
Actually, next_batch() already handles the InMemory variant directly using the zero-copy data[start..end] slice logic and never calls take_batch_from_source() for it. To avoid confusion and make the implementation match the documentation claims, I've replaced the dead-code InMemory arm in take_batch_from_source() with unreachable!().
I've also rebased the branch onto main and resolved all the merge conflicts. Thanks for the review!