CNNs with binary inputs and weights need fixes

[jmitrevs]

Description

As shown by #740, CNNs with binary quantizers don't currently work properly. This PR attempts to fix it.

Type of change

• Bug fix (non-breaking change that fixes an issue)

Tests

A pytest is added. However, it still shows errors in streaming.

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

[jmitrevs]

Before I pushed 7186346 the test results were:

test_binary_cnn.py::test_model2[io_parallel-Vivado] PASSED                                                    [ 16%]
test_binary_cnn.py::test_model2[io_stream-Quartus] FAILED                                                     [ 33%]
test_binary_cnn.py::test_model2[io_stream-Vivado] FAILED                                                      [ 50%]
test_binary_cnn.py::test_model2[io_stream-Vitis] FAILED                                                       [ 66%]
test_binary_cnn.py::test_model2[io_parallel-Quartus] FAILED                                                   [ 83%]
test_binary_cnn.py::test_model2[io_parallel-Vitis] PASSED                                                     [100%]

After, they all fail. So it was not a successful fix.

[jmitrevs]

This is not complete, and the final test has been disabled, but I think the fixes that we have here should be added before we make 0.7.0 RC. The other stuff we can add after the RC is built or later.

The remaining to-dos:

1. Max Pooling does not propagate XnorPrecisionType from the input to the output, so the logic is confused. The output zeros are no longer considered to be -1 but just 0.
2. Conv2d, io_parallel, Resource fails badly with binary inputs--the problem has not been debugged carefully.

If this is accepted, those to-dos should be moved to an issue.

[jmitrevs]

This fixes a few things:

1. cast is called with the correct mult_config in CNN cases in all cases. Previously it was correct in many cases but not all.
2. cast was fixed in the quartus case since you cannot cast ac_fixed to ac_int without calling .to_ac_int(). Maybe one needs to add more templates here depending on whether the input types are ac_fixed or ac_int to cover all the options. (I understand ac_int to ac_fixed casting is allowed.) It's worth looking in more detail since we probably do not cover all cases.
3. The binary versions of batchnorm did not make use of n_scale_bias, which is not the same as n_in when n_filt != -1. They have been updated to match the regular batchnorm. (This was also coped to quartus, where it was missing everywhere.)