Testing/rounding feature on cml trees 4104

[kcelia]

• MSB = 1:

The MSB is set to a value of 1 by default in all our tree-based models. This setting should give the fastest results without decreasing the model's performance. You can review the results of the experiments here.

• Possible approaches to introduce this new feature in our API:

1. Using rounding_bit_padding and auto-rounding:

• To use rounding as truncate, we subtract a term referred to as 'half,' which is calculated as '1 << (lsb - 1).'
• When overflow is detected, an automatic update of the LSB occurs (decremented by 1), which affects the computation of the 'half' value and at the end leading to incorrect results.
• Disabling the overflow check is not recommended

2. Using truncate and auto-truncate:

• The API looks like rounding_bit_padding and auto-rounding
• Not available yet

3. Computing the required bitwidth and LSB via the ONNX Graph: (WORKAROUNG)

• check the overflow
• This method involves checking for overflow and introducing new nodes in the ONNX graph.
• 3.1: We can create new nodes in the ONNX graph. But some tests in 'tests/sklearn/test_sklearn_models.py::test_serialization' may fail due to manual build of 'tree_inference' using fake and non-representative inputs, without considering whether rounding is enabled or not. Because lambda functions can't be properly serialized, we want to limite the use of pickle see this conversation.
• 3.2: We can add a flag lsbs_to_remove(Optional[int]) to existing ONNX nodes

In the current implementation, we have opted for approach 3 + adding a flag to existing nodes in the ONNX graph.
From time to time, we might get an overflow, but the issue problem is not reproducible.
eg: pytest tests/sklearn/test_sklearn_models.py::test_rounding_consistency[6-RandomForestClassifier1] --forcing_random_seed 8576543745180669800 --randomly-dont-reset-seed

The issue with the parallelization shared resources in concrete.fhe.extensions.round_bit_pattern.py is fixed.

closes https://github.com/zama-ai/concrete-ml-internal/issues/4157
ref Add new test for level 2
ref Add a test to check if graphs are identical after and before serialization
ref Control the maximum bitwidth of a circuit with rounding

[jfrery]

Had a quick look. Great job! I have a few comments.

[kcelia]

@RomanBredehoft is it ok for you ?

[jfrery]

I am fine with this PR. Thanks Celia for taking all the comments into account. Let's continue this with truncate now!

[andrei-stoian-zama]

Great work! thanks

[RomanBredehoft]

looks good to me , thanks a lot !

[github-actions]

Coverage passed ✅

Coverage details

---------- coverage: platform linux, python 3.8.18-final-0 -----------
Name    Stmts   Miss  Cover   Missing
-------------------------------------
TOTAL    6539      0   100%

51 files skipped due to complete coverage.