Eliminate dummy batches and init_cuda_buffer.

[stephenroller]

Patch description
Eliminate the dreaded _dummy_batch and _init_cuda_buffer.

Users have complained about the presence of these methods. When debugging, the first batch being a dummy causes some confusion (why am I getting gibberish data?) and the manual implementation of a _dummy_batch is annoying (why is my model breaking on the first pass?)

Adopt Fairseq's newer approach:

• Cache the very first batch we ever seen. If it's invalid, we should be raising an exception anyway. If it's valid, it will be good enough
• Do NOT initialize the cuda buffer. Just keep that batch around
• If we need to recover from an OOM, used the cached dummy batch as the data.

Testing steps
CI.

[klshuster]

thanks for making this change. I'm approving but did have one perhaps pertinent question

[klshuster]

does this unnecessarily use up memory? what if the first batch is huge for example?

[stephenroller]

Could be a worse problem for images. For text, imagine it's a 2x2048x1024 LongTensor (2x for content and label; bs 2048; string length 1024). A long is 8 bytes. So we'll be wasting 32MiB of memory.