Execution context management for TensorRT profiles - #36664

This PR adds a mechanism to handle multiple execution context for a TensorRT engine, which is a step towards enabling dynamic shape mode with multiple profiles.

This PR is not expected to change any existing behavior: inference in implicit batch mode, in explicit batch and in dynamic shapes mode remains the same.

What changes is the engine lookup. Until now the EngineContext object contained a reference to a singe ICudaEngine and a corresponding IExecutionContext. We plan to have one execution context for each profile, therefore we introduce a vector of IExecutionContexts in EngineContext.

Correspondingly the GetEngine function changes: it does not only return the pointer to the EngineContext but also gives the indexd of the IExecutionContext that can be used for inference.

An alternative method for handling multiple profiles would be to keep a single IExecutionContext, and switch profile before we enqueue the new inference task. This could lead to problems with asynchronous inference, because we would need to ensure there are "no enqueued operations pending in the context prior to switching profiles".

Dynamic shape mode is an experimental feature. The advantage of dynamic shapes is that we can use a single engine with multiple profiles to handle input data with different shapes. This saves memory: all input profiles for an engine can share the same weights. Input dimensions that do not influence the number of model parameters can be dynamic (batch size, sequence length, even image dimensions for fully convolutional networks). A single profile can handle a range of input parameters, or we could have multiple profiles each optimized for a certain input shape (e.g. one profile for each expected batch size).

Although this PR implements the mechanism to handle multiple profiles and the corresponding execution contexts, one still needs a method to define these profiles. Such profile generation feature is implemented in PR #36729.