feat: Enable benchmark-mode module inventory/export across all CausalLM architectures

[vbaddi]

WIP: This PR extends enable_benchmark=True support in QEFFAutoModelForCausalLM to all CausalLM models.

What changed

• Added architecture coverage for CausalLM families (lama, gpt_oss including gpt2, codegen, falcon,
  gptj, mistral, mixtral, mpt, phi, phi3, qwen2, starcoder2, granite, olmo2).
• Added mixtral MoE module benchmark support (attention, decoder, moe).
• Added seq_len passthrough to get_benchmark_module_specs(...).
• Added tiny-model benchmark inventory test matrix to validate module dump behavior across all listed tiny
  CausalLM models.
• Kept benchmark behavior gated by enable_benchmark=True; non-benchmark flow remains backward compatible.

Example benchmark output (Llama)

Mode	Module	Type	Prefill ms	Decode ms
Prefill/Decode	Attention	Attention	0.7051	0.5613
Prefill/Decode	Decoder	Decoder	0.7975	0.6924

Input/Output shape section in report

  - attention inputs: {"attention_mask":[1,1,32,128],"hidden_states":[1,32,16],"past_key.0":
    [1,4,128,4],"past_value.0":[1,4,128,4],"position_ids":[1,32]}
  - attention outputs: {"attention_output":[1,32,16],"past_key_RetainedState":
    [1,4,128,4],"past_value_RetainedState":[1,4,128,4]}

Validation

   python -m pytest -q tests/unit_test/benchmarking/test_causal_lm_microbenchmark.py (25 passed)
   python -m pytest -q tests/unit_test/models/test_model_quickcheck.py -n auto (62 passed)

[anujgupt-github]

@vbaddi - can we restructure this as below?
We really need only benchmarks for Attention and FFN (incl expert interactions for MOE models)

We can create Attention and MOE/FFN benchmarks, and use onnx symbols to set fields? Some fields can come from config.json of model card, like dm/dh?

Maybe I didn't fully understand the table you gave above

[vbaddi]

@vbaddi - can we restructure this as below? We really need only benchmarks for Attention and FFN (incl expert interactions for MOE models)

We can create Attention and MOE/FFN benchmarks, and use onnx symbols to set fields? Some fields can come from config.json of model card, like dm/dh?

Maybe I didn't fully understand the table you gave above

Thanks @anujgupt-github. These are all configurable from the config or model card passed, whatever needs to be edited, can either be changed in config or pass that args in .from_pretrained()

The table is basically a dummy inputs running on QAic and providing the numbers for those modules. (/sess.run())