Fix sample_async noise model lifetime and isolation

[huaweil-nv]

Description

Summary

This PR fixes cudaq.sample_async(..., noise_model=...) for local simulators by ensuring the noise model is applied correctly during asynchronous execution and does not leak into subsequent calls.

Root cause

• The async implementation could reference a noise model whose lifetime ended before the queued task executed.
• Noise configuration was not scoped to the async task lifetime, which could cause state pollution.
• Noise set/reset needed to be applied per-QPU (using the provided qpu_id).

Fix

• Extend details::runSamplingAsync to accept an optional noise model and capture it by value inside the async task.
• Set the noise model at the start of the async task and reset it on completion (including exception paths) to prevent state leakage.
• Apply set/reset per-QPU using qpu_id.
• Reject non-empty noise models on remote platforms with a clear error.
• Update Python binding-side remote checks to respect the provided qpu_id.

Tests

Added regression tests in python/tests/builder/test_NoiseModel.py:

• test_sample_async_with_noise
• test_sample_async_noise_isolation

How to repro

import cudaq

cudaq.set_target("density-matrix-cpu")
cudaq.set_random_seed(42)

k = cudaq.make_kernel()
q = k.qalloc()
k.x(q)
k.mz(q)

noise = cudaq.NoiseModel()
noise.add_channel("x", [0], cudaq.DepolarizationChannel(1.0))

# Noise should be visible in async result.
noisy = cudaq.sample_async(k, shots_count=1000, noise_model=noise).get()
print("async noisy:", noisy)

# Subsequent calls without noise must remain clean (no state pollution).
clean = cudaq.sample(k, shots_count=200)
print("after clean:", clean)
assert clean.count("1") == 200

does not always occur

import cudaq, gc

cudaq.set_target("density-matrix-cpu")
cudaq.set_random_seed(42)

k = cudaq.make_kernel()
q = k.qalloc()
k.x(q)
k.mz(q)

def launch_once():
    # Noise model is intentionally scoped to this function.
    noise = cudaq.NoiseModel()
    noise.add_channel("x", [0], cudaq.DepolarizationChannel(1.0))
    fut = cudaq.sample_async(k, shots_count=200, noise_model=noise)
    return fut

futs = [launch_once() for _ in range(200)]
gc.collect()  # Try to encourage destruction of temporaries
for f in futs:
    _ = f.get()
print("done")

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[sacpis]

Overall LGTM. Thanks @huaweil-nv. Let's wait for Thien's review especially for the remote platform part.

[1tnguyen]

LGTM 👍

[1tnguyen]

/ok to test 4714042

Command Bot: Processing...

[github-actions]

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

[khalatepradnya]

Does this address #843?

[huaweil-nv]

Does this address #843?

in Python, cudaq.sample_async(..., noise_model=...) already accepts noise_model, but prior to this change it was not reliable (noise could be ignored and there was a lifetime/UB risk due to a dangling pointer). This PR makes async noise application correct and isolated by capturing the noise model in the async task and setting/resetting it within the task.
in C++, cudaq::sample_async(sample_options{ .noise = ... }) now properly applies the noise model in asynchronous execution (set/reset happens inside the task), and it is qpu_id-correct.

[1tnguyen]

/ok to test 78ce803

Command Bot: Processing...

[github-actions]

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

[huaweil-nv]

/ok to test 64bd8db

Command Bot: Processing...

[khalatepradnya]

/ok to test 9130100

Command Bot: Processing...

[huaweil-nv]

/ok to test 3b24303

Command Bot: Processing...

[github-actions]

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