example/rnn: Batch size must not cause errors in test

[drahnr]

Describe the bug

Currently the rnn mackey glass example under ./juice-examples does not work, if train and test are run with different --batch-size arguments.

To Reproduce

Steps to reproduce the behaviour:

1. Setting up preconditions with ...
2. Run cargo +stable run -p example-rnn-regression --features cuda,native -- train --batch-size=23 foo.capnp
3. Execute cargo +stable run -p example-rnn-regression --features cuda,native -- test --batch-size=1 foo.capnp

Expected behavior

batch_size must not influence success or failure in a sane range.

Please complete the following information:

• System: Fedora 33 / Windows x64
• Version: *
• Rust: rustc 1.51.0 (2fd73fabe 2021-03-23)
• Environment: CI env
• Backends (if relevant):
  • opencl: CI env
  • cuda: CI env

Additional context

[drahnr]

As @lissahyacinth pointed out in direct comms, it's not related to RNN but to the Reshape layer which serializes all dimensions into the .capnp file, which in turn is restored for test runs of the example - which might have a different batch dimension size compared to the train run.