Roll-forward:

PR #24081: Split convolution invocation into preparation and actual invocation

Please approve this CL. It will be submitted automatically, and its GitHub pull request will be marked as merged.

Imported from GitHub PR #24081

- split DoConvolve into:

PrepareForConvolution
DoConvolve

- split DoConvolveBackwardData into:

PrepareForConvolutionBackwardData
DoConvolveBackwardData

- split DoConvolveBackwardFilter into:

PrepareForConvolutionBackwardFilter
DoConvolveBackwardFilter

PrepareForConvolutionXXX would allocate scratch memory.
DoConolveXXX would invoke actual convolution algorithms.

Implement forward convoution, backward input convolution, backward filter
convolution on CUDA path.

Copybara import of the project:

  - fe9a9dbb Split convolution invocation into preparation and actual ... by Wen-Heng (Jack) Chung <whchung@gmail.com>
  - baeda0d8b5da58a97d28ddce264c80c58d937699 Merge fe9a9dbb into bbe33... by Wen-Heng (Jack) Chung <whchung@gmail.com>

PiperOrigin-RevId: 228568812