✨[Feature] Support INT64 inputs at the graph input level

[peri044]

Is your feature request related to a problem? Please describe.
Currently if the graph is given int64 inputs, the compilation fails with unsupported type.

Describe the solution you'd like
Solution we are expecting : Cast INT64 inputs to INT32 and issue a warning to users about this instead of breaking compilation.

Describe alternatives you've considered

Additional context

[gs-olive]

This feature may warrant a multi-phase detailed RFC, as it relates closely to Issue #1346, and the current (Draft) implementation can cause issues for users requiring aten::scatter support or other Torch-only Long-only operations. The linked PR #1551 implements autocasting in a way which is efficient (no-copying) but can break existing models due to issues with Int64-required operations, since it in-place casts input tensors for the duration of inference. I think additional design work is needed to track data types through different engines and implement an efficient casting system.

RFC: #1553

[Christina-Young-NVIDIA]

Feature split between MVP & some optional phases (detailed in RFC), trying to complete MVP for this sprint.