[GPU] Block the fc horizontal fusion for weights that are not constants

[intbf]

Details:

Found when using weight sharing with llms. In that case weights are converted into inputs/parameters, and the fusion might not work as expected. With inputs/parameters the constant folding won't happen, so the fusion will be invoked on every inference. Plus in some cases the concat operator might not be enabled for all weight types (like int4)

Tickets:

CVS-189072

AI Assistance:

• AI assistance used: yes
• do the code changes + unit tests

[Copilot]

Pull request overview

This PR updates the Intel GPU plugin’s FullyConnectedHorizontalFusion transformation to avoid fusing horizontal FC patterns when FC weights are not compile-time constants, addressing weight-sharing scenarios where weights become runtime Parameters and would otherwise force an unfused Concat into GPU program build (and/or cause per-inference overhead).

Changes:

• Add a guard in the horizontal-fusion pattern predicate to require FC weight inputs to be constant (or constant-derived) before allowing fusion.
• Add unit tests covering weights-as-Parameters and Convert(Parameter) weight forms to ensure fusion is blocked in these cases.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/plugins/intel_gpu/src/plugin/transformations/fc_horizontal_fusion.cpp	Blocks horizontal FC fusion when weight input is not constant, preventing unfused runtime Concat and avoiding per-inference concatenation overhead.
src/plugins/intel_gpu/tests/unit/transformations/horizontal_fc_fusion_test.cpp	Adds regression tests ensuring no fusion occurs when weights are provided as Parameter (including Convert(Parameter)).

[e-ddykim]

Looks good to me