Rework ORTQuantizer & ORTOptimizer to easier work with already converted `onnx` checkpoints and add support for seq2seq models (RFC)

[philschmid]

ORTQuantizer

• [x] remove hard transformers dependency
• [x] rework from_pretrained to include from_transformers, ORTModelForXXX or a path to a model.onnx file
• [x] add file_name to from_pretrained
• [x] add better docstrings with example
• [x] align parameters for method, e.g. export
• [x] add more tests for creating ORTQuantizer

ORTOptimizer