Fix onnx calibration state restoration

[Sh0rck_Wang (WangXukang-cypher)]

Fixes #4114

Summary

The ONNX compute_encodings context previously left quantizers in a partially calibrated state if the calibration callback or encoding finalization raised an exception.

This change makes calibration exception-safe by:

• Preserving enabled quantizers' existing encodings and operation modes before calibration.
• Restoring the preserved encodings and operation modes when calibration exits with an exception.
• Re-raising the original exception after restoring quantizer state.
• Making nested compute_encodings contexts reuse the outer calibration lifecycle so that an inner context does not reset statistics or finalize encodings prematurely.
• Adding regression tests for exceptional exits and nested calibration contexts.

The exceptional-exit regression test verifies that:

• Existing encoding values are preserved.
• Quantizer operation modes are restored.
• Quantized forward outputs remain unchanged after a failed calibration attempt.

Test plan

The changes were tested in a Python 3.10 CPU environment with an editable source build of AIMET ONNX.

• Run the new exceptional-exit and nested-context regression tests.
• Run the existing normal compute_encodings tests.
• Run all compute_encodings-related tests in test_quantsim.py.
• Run repository pre-commit checks on the modified files.
• Run IDE lint checks on the modified files.

Results

4 passed