Describe the feature request

It would be great to have the option to provide pre-optimised TensorRT engine plans to ORT.

Describe scenario use case

Using TensorRT in standalone, e.g. trtexec, allows for much more precise control over engine compilation.

Currently, only a very small amount of TRT features is accessible through the ORT API, but most importantly, there is no way to ship precompiled engines with an application; Engine compilations need to be done on the user machine, resulting in a bad user experience.
Also, how (TRT) caching is handled internally is not very transparent, in my opinion, and might lead to unnecessary engine recompilations.