Align locality-domain MoE quantization detection with Linear planning

Summary

Align quantization detection in LocalityDomainExecutionPlanner.plan_moe with LocalityDomainExecutionPlanner.plan_linear.

Rationale

plan_moe currently reads quant_config.quant_mode and calls has_any_quant() without excluding KV-cache-only quantization. plan_linear reads the intended layer quantization source and excludes KV-cache-only quantization. The two planners can therefore classify equivalent configurations differently. In particular, a KV-cache-quantized BF16 MoE can be rejected as unsupported instead of using the BF16 locality-domain MoE path.

Affected area

• tensorrt_llm/_torch/locality_domain/policy.py
• LocalityDomainExecutionPlanner.plan_moe

Required changes

• Confirm and use the intended MoE layer quantization attribute.
• Use the same quantization source for NVFP4 detection and general quantization detection.
• Exclude KV-cache-only quantization when determining whether weight quantization is present.
• Classify KV-cache-quantized BF16 MoE configurations as BF16 when all other BF16 requirements are met.

Acceptance criteria

• plan_moe and plan_linear apply consistent effective-weight quantization semantics.
• A KV-cache-only-quantized BF16 MoE does not reach the unsupported-quantization branch solely because of KV-cache quantization.
• Unit coverage verifies the expected classification and disabled reasons.

Backlinks

• Merge-back PR: https://github.com/NVIDIA/TensorRT-LLM/pull/18369
• Review discussion: https://github.com/NVIDIA/TensorRT-LLM/pull/18369#discussion_r3881378822

Requested by: @farazkh80