[RFC] Drop legacy cuDNN releases support in MXNet 2.0

[mk-61]

In v8 cuDNN introduced new API. It enables more generalized support for heuristics, numerical properties of compute engines and engine knobs. It is substantially different to the legacy one, so new convolution op cache has been implemented in MXNet container and the old one has been removed. In order to upstream new API support we either need to maintain 2 convolution op caches, or drop the legacy one.

Maintaining 2 op caches would keep legacy cuDNN releases supported, but would make the code messier.

Secondly, it might be less painful to make this move now rather than later, with more customers moved to 2.0.

Currently this change affects only convolution and deconvolution ops, but more operations are expected to be ported to the new API, including fused operations.

[szha]

+1