Add a eager mode for PyTorch/XLA. We already have an eager mode for debug which is UseEagerDebugMode that's controlled by a env var. This was contributed by the AWS folks(cc @aws-rhsoln @amithrm @jeffhataws). I decided to make a offical top level api for it.

In this pr, I

1. add api use_eager_mode along with all the pybinds
2. make sure that PT_XLA_DEBUG and PT_XLA_DEBUG_LEVEL does not print analysis for eager execution
3. add new metrics for eager mode compilation and eager mode executions

For future pr

1. increase the maximum number of queued task (for eager mode)
2. add api/decorator to compile a specified function in the eager mode.
3. add docs
4. test eager + SPMD

For a 2 layer decoder only model, eager mode can acheive ~40% of throuput compared to the full compiled mode.