Bring back IFB default to TRT LLM models and bump to 24.01

[rmccorm4]

• Brings back default IFB mode for TRT LLM since input_ids issue is fixed in 24.01 trtllm backend
• Bumps dockerfile version to 24.01
• Adds --log-verbose=1 to Triton server start when -v verbose flag is used
• Sets default profiling mode back to "online"

[rmccorm4]

Some of our server messages are logged at verbose level 2 (AFAIK nothing higher than that). Is there any value in bumping the default level to 2 since the user is explicitly asking for verbose logging?

Great question! Totally possible to update in the future if needed, but personally I find 2+ to be a bit overkill for most cases. It currently will print constantly when using any sequence batching models, even if no inference is going on, for example - and goes a bit crazy when caching is enabled too. Happy to update later if needed.