Enable Dynamo KV-blocking Loop export for kv_headpar

[quic-abhamidi]

Integrate a Dynamo-only torch.while_loop implementation for blocked_kv_attention_forward_headpar_offline so KV-blocked decode can export a structured ONNX Loop instead of unrolling the Python for-loop per KV block.

Key changes:

• Enable the KV Loop path by default for blocking_mode=kv_headpar when dynamo=True.
• Preserve the legacy Python-loop path for dynamo=False and allow explicit opt-out with use_kv_loop_op=False.
• Add StaticLoopInputsTransform to rewrite ONNX Loop trip count and initial condition to static constants required by QAIC compiler unrolling.
• Add InlineTorchSubgraphFunctionsTransform for dynamo subfunction export, inlining torch-generated subgraph functions that contain Loop nodes so QAIC does not see Loop control inputs inside FunctionProto bodies.
• Update Dynamo causal LM examples and README with KV-blocking CLI support.
• Add Dynamo/unit tests covering Loop export, config defaults, static Loop
  input rewriting, and function-body Loop handling.

Verified:

• tests/dynamo/test_kv_blocking_loop.py: 4 passed
• tests/unit_test/transforms/test_blocking_transform.py::TestDynamoKVLoopBlockingConfig: 3 passed
• Qwen2-1.5B kv_headpar Dynamo compile succeeds with and without
  --enable-onnx-subfunctions.