Faster data unpacking (Diffferent to #1291)

[klassen9]

Here is a faster implementation of the data unpacking (src/finn/util/data_packing.py::packed_bytearray_to_finnpy). This implementation is different to the one seen in #1291.

While being very efficient mdanilows variant suffers from weaknesses, such at not supporting SIMD>1 and not supporting some data types such as fixed and floating point.

This PR addresses these problems and the performance problems of the current implementation by adding a unique unpacking for different datatype categories.

Furthermore, I removed the inferring of the output_shape, as it is ambiguous: E.g. a Byte can store 1 to 4 UINT2 numbers.

I ran a test comparing the current implementation to my variant:

For the first test I assumed an input of shape (10, 32, 32, 8, 1) of different datatype, which is packed into a byte array. Then I unpacked it with both variants. Here are the speedups:

For the second test I assumed an input of shape (10, 8, 1). One can see that the speedup is decreasing, but still is substantial:

[auphelia]

As part of this PR, I've also moved the get_driver_shapes function into finn.util.basic because the data_packing.py is used as part of the deployment folder that gets copied to the device, get_driver_shapes is not required in there but has additional dependencies (like ModelWrapper and onnx). Thanks @mdanilow for making me aware of this!

[auphelia]

Thank you, @klassen9!