Backend version update

[maljoras]

Related issues

closes #346, closes #291, closes #290

Description

Many updates and fixes in particular to the RPUCuda backend, including

• Supporting analog tiles moving from GPU to CPU
• non-blocking option for tile computation in CUDA
• Input range learning for inference
• bump pylint / mypy / pycodestyle versions
• Added support for N-degree polynomial weight noise models for hardware-aware training for analog inference .
• Reset method for tiles on updated device
• Many more non-linearties for the analog mat-vec including multiple passes (positive / negative)
• Hidden weight write noise fixes
• Added a remap_weights for AnalogSequential and AnalogModuleBase level.
• Improved memory management of CUDA tiles
• PowStepReference and SoftBoundsReference device, with option to subtract symmetry point (see Tool/tutorial for computing the symmetry points of a given device and setting up a TransferCompound  #291)
• Fitting utility (Tool for fitting device data to models in AIHWKIT #290)

Details

Input range learning

Learning of the input clipping bound.

• Gradient depends on clipping of input as well as output
• Decay
• adjustable parameters in InputRangeParameters

Yields very good results (see example 06)

non blocking

• CudaTiles can be made no blocking by setting tile.non_blocking = True

[kaoutar55]

Thanks @maljoras! can we merge this before the 0.7.0 release?