Enable FP32 support for thresholding

[SpiritSeeker]

• Infer thresholding HW layer only when the input and thresholds are both integers or both FP32.
• Enable use of FPARG in thresholding RTL for FP32 comparisons
• Add tests for FP32 thresholding

[auphelia]

Thank you @SpiritSeeker ! Looks great :)

[fpjentzsch]

@SpiritSeeker @auphelia On our end this broke 3 tests. Please double check and see if you need these fixes on your side as well.

1) test_convert_to_hw_layers_cnv_w1a1[False]

The input quantizer is now unexpectedly converted to thresholding_hls. I fixed this here: eki-project@9d4c654

Note that 1) an additional AbsorbConsecutiveTransposes() was needed and 2) InferConvInpGen() had to be moved to not leave 2x Im2Col in front of the pooling layers (this adjustment was probably missed during the recent consolidation of the pooling operators).

I wonder if this will have negative/unexpected implications for other models used in examples/regression tests...

2) test_fifosizing_linear

Error encountered in minimize_accumulator_width:

Exception: Could not find a suitable int datatype for -3.4028234663852886e+38

I fixed this by enforcing FP32 datatype for thresholds in case of FP32 inputs: eki-project@1e99b39

3) test_split_large_fifos

As with the cnv, the model now has 1 additional thresholding layer + DWC, leading to 2 additional FIFOs that were not accounted for. Fixed here: eki-project@d72df88

[fpjentzsch]

For some models, like the KWS example, this also uncovers streamlining problems that we should probably address with proper layout conversion in the front-end and more robust handling of Transpose/Reshape/Flatten operators. See eki-project#101