[GPU] Enable DynamicQuantize sharing to reduce redundant quantization

[hyunback]

Problem

When DynamicQuantize (DQ) nodes are inserted before FullyConnected operations during GPU plugin graph optimization, identical DQ nodes with the same input are created independently for each FC consumer. This results in redundant quantization computation and increased graph complexity.

Solution

Register SharedOpOptimization after DynamicQuantize insertion to deduplicate identical DQ nodes sharing the same input. Relax single-user assertion in dynamic_quantize graph node to allow shared consumers.

Gemma-4-26B-A4B-it

• 156 DQ nodes -> 121 DQ nodes (35 eliminated, ~22% reduction)
• gate+up FC sharing: 30 pairs across all 30 layers (gate_proj and up_proj share the same FFN input after RMS normalization)
• Q+K FC sharing: 5 pairs in sliding window layers (L5, L11, L17, L23, L29 have separate Q/K FCs sharing the same attention input)

SD3.5 large

Model	DQ Sharing OFF	DQ Sharing ON	Reduced
0 (text encoder)	98	98	0
1 (text encoder)	258	258	0
2	240	192	48
3 (transformer)	768	616	152
4 (decoder)	8	4	4
Total	1372	1168	204

Tickets:

• CVS-166834

AI Assistance:

• *AI assistance used: yes
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).