[HybridAttention] Support past key values in attention_mask for stateful inference

[apaniukov]

What does this PR do?

GraniteMoeHybrid crashes with a shape mismatch error:

RuntimeError: Exception from src/inference/src/cpp/infer_request.cpp:224:
Exception from src/plugins/intel_cpu/src/node.cpp:787:
[CPU] Add node with name '__module.model/aten::add/Add_1' Check 'input_shape[j] == 1' failed at src/plugins/intel_cpu/src/shape_inference/custom/eltwise.cpp:52:
Eltwise shape infer input shapes dim index: 3 mismatch

The RuntimeError at the Add node that combines the causal mask with the position-expanded attention_mask:

• port 0 (causal mask): [batch, 1, seq_len, seq_len]
• port 1 (attn mask): [batch, 1, 1, total_seq_len]
  where seq_len != total_seq_len during 2nd+ generation steps when seq_len > 1 (e.g. second prefill stage).

The issue caused by patched _update_causal_mask method. Hybrid models that doesn't patch it (Qwen3-Next and LFM2) work fine.

Fix:

• Update inputs property to use "past_sequence_length + sequence_length" for attention_mask for -with-past conversion
• [unrelated] remove unnecessary triu operation

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[rkazants]

can you please add test to reproduce this case?

[apaniukov]

can you please add test to reproduce this case?

Done.