Lookahead decoding and multiple profiles support

Hi,

After #3137 was solved, lookahead decoding with multimodal inputs seems to work correctly, but only when multiple_profiles==disabled. When it's enabled I get the following error:

[TensorRT-LLM][ERROR] ICudaEngine::getIOTensorName: Error Code 3: API Usage Error (Parameter check failed, condition: index >= 0 && index < getNbIOTensors(). Invalid IO Tensor Index: 29. Valid range of IO Tensor Indexis [0, 26). Upper bound of IO Tensor Index can be queried with ICudaEngine::getNbIOTensors.)
[TensorRT-LLM][ERROR] IExecutionContext::inferShapes: Error Code 3: API Usage Error (Parameter check failed, condition: validDimension. Supplied binding dimensions [1] for tensor (getIOTensorName returned nullptr) exceed min..max range for profile 1. Valid range is [64]..[128].)
[TensorRT-LLM][ERROR] Encountered an error in forwardAsync function: Invalid input shape (/home/jenkins/agent/workspace/LLM/release-0.18/L0_Test-x86_64/tensorrt_llm/cpp/tensorrt_llm/runtime/tllmRuntime.cpp:574)
1       0x7fd1e98c82b7 /home/maximilianolevi/.cache/pypoetry/virtualenvs/tensorrt-inference-8MUMp6os-py3.10/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(+0x6e52b7) [0x7fd1e98c82b7]
2       0x7fd1ea5ac458 tensorrt_llm::batch_manager::TrtGptModelInflightBatching::prepareBuffers(std::vector<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > > const&, std::vector<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > > const&, int) + 184
3       0x7fd1ea5b59a6 tensorrt_llm::batch_manager::TrtGptModelInflightBatching::executeStep(std::vector<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > > const&, std::vector<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > > const&, int) + 1510
4       0x7fd1ea5b628f tensorrt_llm::batch_manager::TrtGptModelInflightBatching::executeBatch(tensorrt_llm::batch_manager::ScheduledRequests const&) + 223
5       0x7fd1ea5bfb7a tensorrt_llm::batch_manager::TrtGptModelInflightBatching::forwardAsync(std::list<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > > const&) + 1802
6       0x7fd1ea64c775 tensorrt_llm::executor::Executor::Impl::forwardAsync(std::list<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest>, std::allocator<std::shared_ptr<tensorrt_llm::batch_manager::LlmRequest> > >&) + 437
7       0x7fd1ea6584e6 tensorrt_llm::executor::Executor::Impl::executionLoop() + 1206
8       0x7fd4f2a295c0 /home/maximilianolevi/.cache/pypoetry/virtualenvs/tensorrt-inference-8MUMp6os-py3.10/lib/python3.10/site-packages/torch/lib/libtorch.so(+0x145c0) [0x7fd4f2a295c0]
9       0x7fd4fddedea7 /lib/x86_64-linux-gnu/libpthread.so.0(+0x7ea7) [0x7fd4fddedea7]
10      0x7fd4fdf03acf clone + 63

To reproduce:

• Enable lookahead decoding
• Compile with multiple_profiles enabled
• Request has multimodal input (e.g. just custom prompt table with fake vocabulary)
• batch size > 1
• Inflight fused batching is enabled
• Model is Llama 8B.

Reproduce the failure with lookahead decoding, multiple_profiles, multimodal input, batch size greater than 1, inflight fused batching, and the Llama 8B model. Start in tllmRuntime.cpp around line 574 and inspect TrtGptModelInflightBatching::prepareBuffers; done means this configuration completes without invalid tensor-index or input-shape errors.