Copy behaviour of e. g. `free` and `fftw_plan_destroy`, where deleting null pointer does nothing instead of segmentation fault.

Add NULL pointer check to deleteVkFFT

…real transform per axis (as uint)

-Solves #151

…fn/pyvkfft#33)

-Allows to change descriptors inside one command buffer
-Needs to be enabled by the user before  vkCreateDevice and set with usePushDescriptors flag in configuration of VkFFT
-there is a warning in validation layer that needs to be investigated "vkCreateDevice: pCreateInfo->pNext chain includes a structure with unexpected VkStructureType VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR"

-improved glslang management in CMake
-replaced the relative path to the glslang_c_interface with the full path

-fixed missing line that was lost somewhere during v1.3.0 reorganization of read/write module

-enabled through singleKernelMultipleBatches parameter
-kernel batching is controlled through coordinateFeatures
-number of input/output systems is controlled through numberBatches
-sample 53 shows the usage of this option

-Real and imaginary parts of a number can now be supplied from two buffers with their own respective offsets (so both buffers can actually be just one with an offset for I).
-Any buffer array pointer can be provided in such form (as two buffers) - input, output, buffer, kernel, temp. Need to set the respective bufferNum and bufferSeparateComplexComponents value.
-Updated samples 0, 50 and 11 (for now) to demonstrate how this feature works.
-Bugfixes.

-Improves performance for big sequences
-Optimized radix kernels to use fewer instructions
-Use __ldg in CUDA and HIP

-Fixed some warnings

…rialization

-Option to use temp buffer to automatically pad every intermediate memory access to break power of 2 strides. Can boost performance up to 2x in some cases. Enabled in all APIs for AMD (can be force-overwritten by user).

-Force the callback version of R2C if zeropadding is enabled. The even decomposition version is hard to make work for proper zeropadding, as it is read as complex numbers and won't support odd zeropadding lengths trivially.