Follow-up from AIPCC-11499 (CPU .cpu tag support). The vllm.py tag matcher in package_plugins currently groups TPU with CUDA (both use accelerator="cuda"). As we start cutting variant-specific tags (e.g. .tpu, .neuron, .spyre), the matcher needs to support per-variant accelerator suffixes so each variant can resolve its own tags.

Current tag suffixes in use in nm-vllm-ent:

• .cpu (v0.14.1+rhai1.cpu, v0.13.0+rhai6.cpu)
• .cuda (v0.11.2+rhai0.cuda, v0.11.2rc1+rhai0.cuda)
• .rocm (v0.11.2rc1+rhai0.rocm)
• .neuron (v0.11.0+rhai0.neuron)

Not yet in use but anticipated:

• .tpu
• .spyre

The _create_vllm_matcher_with_legacy function should be updated to give each variant its own accelerator value so the ADR-114 matcher can resolve variant-specific tags independently.

Related to AIPCC-11499 and INFERENG-4612.