Optimize Decoder Pipeline Model Execution

[baijumeswani]

#729 introduced support to execute a pipeline of ort sessions sequentially as defined by the config file.

This pull-request builds on top by:

• Fixing a memory leak bug where the ortvalues stored in the ortvalue_pool_ were raw pointers and were not released.
• Clears outputs from pipeline state that was only run on prompt. Before this pull-request, those ortvalues lived in the ortvalue_store_ forever using unnecessary memory.
• Enable registering the managed inputs as pipeline model outputs.
• Sharing an allocator across ort sessions.

Co-contributor: @edgchen1 who helped identify the bug(s).

[baijumeswani]

Thank you for the review. :)