Enables PA for arm64

[ilya-lavrenov]

See openvinotoolkit/openvino#29219

CVS-164838

[Copilot]

Pull Request Overview

This PR enables paged attention (PA) support on ARM64, broadening the scope of architectures that can leverage this functionality.

• Removed the forward declaration for with_cpu_sve() since its runtime support is no longer conditionally required.
• Modified preprocessor conditions to enable PA on ARM64 along with x86_64.
• Updated error messages to consistently reflect that PA is supported on x86_64 and ARM64 platforms only.

Comments suppressed due to low confidence (3)

src/cpp/src/llm/pipeline.cpp:50

• The updated error message for the continuous batching backend now correctly reflects support on ARM64. Ensure that the relevant documentation is updated accordingly.

OPENVINO_THROW("Continuous batching backend requires PagedAttention operation support, which is available on x86_64 or ARM64 platforms only");

src/cpp/src/llm/pipeline.cpp:57

• The error message for speculative decoding now correctly indicates support on both x86_64 and ARM64. Confirm that this behavior aligns with the expected backend logic for ARM64.

OPENVINO_THROW("Speculative decoding requires PagedAttention operation support, which is available on x86_64 or ARM64 platforms only");

src/cpp/src/llm/pipeline.cpp:64

• The error message for prompt lookup decoding is updated to reflect ARM64 support. Verify that there is no implicit dependency on SVE requirements within the ARM64 code path.

OPENVINO_THROW("Prompt lookup decoding requires PagedAttention operation support, which is available on x86_64 or ARM64 platforms only");