Read number of workspaces from argument in recurrent_network_op (#23272)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23272
We see significant performance improvements by limiting concurrency
at caffe2 level on mobile. This diff enables setting the number of caffe2
workspaces used during rnn inference.
Reviewed By: akyrola
Differential Revision: D16448611
fbshipit-source-id: 28abaddb4ea60bacb084ceb28cb7a4d1e67ccc17