feat(0506): Add optional KV-cache buffer-name prefix for vLLM disaggr…

[vbaddi]

Summary

• model classes (QEFFAutoModelForCausalLM, single-QPC VLM, dual-QPC VLM).
• When provided, injects the token as an infix into KV-cache retained-state buffer names:
  past_key.{i}_RetainedState → past_key.{i}_<prefix>_RetainedState, enabling vLLM to
  regex-select LLM KV buffers for disaggregated device-to-device transfer.
• Vision/multimodal retained buffers are never renamed. Default behaviour is unchanged.

Motivation

The vLLM team integrating on this QEff branch needs a stable regex handle on LLM KV buffers
(distinct from vision_embeds_RetainedState etc.) for disaggregated KV transfer between devices.

Test plan

• python -m pytest -q tests/unit_test/models/test_model_quickcheck.py -n auto
• All 121 existing tests still pass
• 18 new unit tests cover: helper correctness, export prefix on CausalLM, default unchanged,
  hash-dir collision prevention, compile custom_io consistency, VLM lang-only scope, and input
  validation

Usage

## CausalLM
  model = QEFFAutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
  model.compile(prefill_seq_len=32, ctx_len=4096, num_cores=16, kv_cache_prefix="VLLM")
  # KV buffers: past_key.0_VLLM / past_key.0_VLLM_RetainedState
  
## VLM — dual QPC (kv_offload=True)
  model = QEFFAutoModelForImageTextToText.from_pretrained("meta-llama/Llama-3.2-11B-Vision-Instruct", kv_offload=True)
  model.compile(prefill_seq_len=32, ctx_len=4096, img_size=560, num_cores=16, kv_cache_prefix="VLLM")

# Without the flag — behaviour identical to today, no name changes.
  model.compile(prefill_seq_len=32, ctx_len=4096, num_cores=16)