Update TensorRT-LLM

[Shixiaowei02]

• Model Support
  • Support Qwen1.5 MoE A2.7B
  • Support Phi 3 vision multimodal
• Features
  • Encoder-Decoder C++ Runtime TP Support
  • Explicit draft tokens inflight batching
  • Support local file for calibration
    • Thanks to the contribution from @DreamGenX: Support custom calibration datasets #1762
  • Add batched logits post processor
  • Add Hopper qgmma kernel to XQA JIT codepath
  • MoE enable TP+EP
  • Add lookahead decoding layer
• API
  • [BREAKING CHANGE] Setup buffers for explicit draft tokens decoding
  • [BREAKING CHANGE] Replace all occurrences of max_output_len with max_seq_len
    • This involves trtllm-build and benchmark related parameters
  • [BREAKING CHANGE] Remove GptSession Python bindings
  • [BREAKING CHANGE] Add runtime max batch size to gptManagerBenchmark
  • Support remaining executor API options in HLAPI
  • Support get_stats and aget_stats in HL Executor while using multi-gpu
  • Add iterLatencyMilliSec to stats and iteration log
• Bug fixes
  • Can't convert-checkpoint Mistral 7B v0.3
    • Thanks to the contribution from @Ace-RR: Can't convert-checkpoint Mistral 7B v0.3: safetensors_rust.SafetensorError: File does not contain tensor model.embed_tokens.weight #1732
  • Inflight batching for fp8 Llama and Mixtral is broken
    • Thanks to the contribution from @bprus: Inflight batching for fp8 Llama and Mixtral is broken #1738
  • quantize.py fails to export important data to config.json
    • Thanks to the contribution from @janpetrov: quantize.py fails to export important data to config.json (eg rotary scaling) #1676
  • Refactor the dynamic decoder params
  • Fix long runtime for MOE models when using FAST_BUILD
  • Enhance ITensor::slice to extreme cases
  • HLAPI exits gracefully on exceptions
  • NaN appears in the result under the one shot all reduce strategy
  • Cache ncclComm_t as weak_ptr and wrap it as shared_ptr to avoid accidentally destroyed
• Memory optimization
  • Support stream reader to reduce peak memory when using weight streaming
• Benchmark
• Performance
  • Optimize the build time when XQA JIT is enabled
  • Reduce number of stream when using fused decoder
• Infra
• Documentation
  • Update documents about GEMM plugins
  • Polish enc-dec readme to reflect recent changes
  • Update Mixtral example docs to include Mixtral-8x22B instructions
  • Simplify recurrent gemma README