feat: QNN Multi Chunk Execution in New Frontend

[oreomaker]

Multi Chunk Execution

Tokenizer: Create a new tokenizePaddingByChunk in SmolLMTokenizer, which will takes input and padding to nearest multiplication of chunk_size

auto [real_seq_length, input_tensor] = tokenizer.tokenizePaddingByChunk(input_str, chunk_size, config.vocab_size);

Module Static States: Add Module::isMultiChunkPrefilling and Module::isFirstChunk to record the multi chunk execution

Module Execution: Add a new tensor_status of TENSOR_UNDEFINED, which is used in QNN chunk execution. If Module::isMultiChunkPrefilling is true, the QNN modules will not reshape & setUp in following chunks, while CPU modules still reshape & setUp

if (Tensor::tensor_status == TENSOR_STATIC_INIT && device_ != MLLM_CPU) { // backend specific module reshape & setup
                if (Module::isMultiChunkPrefilling && !Module::isFirstChunk) {        // set to TENSOR_UNDEFINED and SKIP executing qnn layers
                    Tensor::tensor_status = TENSOR_UNDEFINED;
                    auto outputs =  Forward(inputs, anyArgs);
                    Tensor::tensor_status = TENSOR_STATIC_INIT;
                    return outputs;
                }
...

TODO

Multi round input still output weird results, which may be caused by stateful OPs like KVCache, RoPE and CasaulMask