[QDP] Add CUDA kernel tests for basis encoding

[tongyu0924]

Changes

• Test

Why

basis.cu had no test coverage.

How

Add tests/basis_encode.rs covering:

• Single-sample encode: first, middle, last index (f64 + f32)
• Batch encode: multiple samples with different indices (f64 + f32)
• Error rejection: out-of-range index, zero state_len, zero num_samples
• Non-Linux dummy stub

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[ryankert01]

Can we have multiple test cases and hard testcase that test the limit and edge case of the code?

[400Ping]

Please fix pre-commit

[400Ping]

Thanks for adding kernel-level test coverage for basis encoding. I think this PR should be tightened in a few places:

1. Please reduce the duplicated test setup/assertion logic.
   Most tests repeat CUDA device creation, allocation, launch, copy-back, and basis-state assertions. A small helper such as cuda_device_or_skip(), assert_basis_state_f64(), assert_basis_state_f32(), and table-driven cases for first/middle/last would make this much easier to maintain.

2. Please add or clarify the batch-kernel edge cases.
   The batch kernels rely on state_len == 2^num_qubits because they compute:

   • sample_idx = global_idx >> num_qubits
   • element_idx = global_idx & (state_len - 1)

   Right now the launcher only rejects zero num_samples / state_len. If the intended contract is that callers must validate state_len, num_qubits, and basis-index bounds before calling the launcher, please document that in the test or PR description. Otherwise, please add validation/tests for:

   • batch out-of-range basis index
   • state_len / num_qubits mismatch
   • f32 batch zero state_len

3. Please assert the imaginary part in the f32 batch test too.
   test_basis_encode_batch_f32_basic currently checks only actual.x, while the f64 batch test checks both real and imaginary values.

[rich7420]

@tongyu0924 thanks for the patch!

nits: the state buffers are created with alloc_zeros, so the "all other elements should be 0" assertions pass even if the kernel never writes those slots — they're zero to begin with. Since basis encoding's contract is to overwrite the entire state vector, pre-filling the buffer with a sentinel (e.g. 1.0) before launch would actually verify the kernel zeroes them. Matters most for the batch kernels, where a grid-stride indexing bug could silently skip elements.