Revamp MultiThreshold

[maltanar]

• Clean reimplementation of the core multithresholding implementation based solely on numpy broadcasting and vectorized ops, no more loops. Much faster and capable of supporting arbitrary dimensionality/data layout.
• Allow core multithresholding op to optionally do channels-last conversion where needed, simplifying the MultiThreshold op wrapper.
• Bring back performance benchmark in test_multithreshold.py
• Expands test cases for the core multithresholding implementation to cover more combinations of dimensions and tensor shapes

Example run from performance benchmark:

input size (128, 1024, 1) with 254 global (not per-channel) thresholds

before:
Runtime non-vectorized:  31.17622971534729 s
Runtime vectorized:  128.39458513259888 s
Speed-up:  0.24281576737173294

after:
Runtime non-vectorized:  31.00561213493347 s
Runtime vectorized:  0.057237863540649414 s
Speed-up:  541.6975794862396