fix(optim): support like factories for placeholder contexts

[Zhaoxian-Wu]

Why

AnalogContext exposes placeholder data by default so public tensor operations cannot accidentally read meaningless backing weights. Tensor *_like factories only need metadata such as shape, dtype, and device, and should be valid in this mode. Before this change, they were rejected as value-reading operations.

This is related to the placeholder-data feature introduced by IBM/aihwkit#765, and completes its compatibility with standard PyTorch *_like factories.

Failure case

model = AnalogLinear(1, 1, bias=False, rpu_config=SingleRPUConfig(
    device=ConstantStepDevice()
)).to("cuda")
ctx = next(model.parameters())
torch.zeros_like(ctx)

Before this fix, the final line raised:

RuntimeError: AnalogContext data is in placeholder mode, so operation 'zeros_like' cannot read weight values.

This prevented metadata-only tensor construction during CUDA model setup and other generic PyTorch code paths.

What changed

• Allow empty_like, full_like, ones_like, rand_like, randint_like, randn_like, and zeros_like in placeholder mode.
• Keep the placeholder backing storage unread while allowing PyTorch to allocate the result from metadata.
• Add CPU coverage for all seven public *_like factories and a CUDA regression test for zeros_like.

[Zhaoxian-Wu]

Hi @PabloCarmona,
All the test passed in my local environment. However, the test failed intermittently (test-and-lint / Test 3.12) because two random noisy outputs happened to quantize to the same value. Could you please review this stochastic test case?

Possible solution: tests/test_simulator_tiles.py:686 compares only one randomly selected noisy-output pair, so it can falsely fail when both samples quantize to the same value.

[PabloCarmona]

Thanks @Zhaoxian-Wu ! I will take a look at the PR and what are you talking about the tests for the random noisy value, we are aware of it, we will take a look again.

Thanks again!

[PabloCarmona]

@Zhaoxian-Wu could you sync up this branch with the master branch? We are planning to launch a new release very soon. Thanks again!