batches: improve failure behaviour when uploading changeset and batch specs

[LawnGnome]

One problem that we've now seen from multiple users is that uploading changeset and batch specs after batch change execution is fragile — since there are n+1 uploads, any load issue on the Sourcegraph instance or network blip beyond what a single HTTP request can handle is problematic.

We should either implement some basic retry/backoff logic in src, or provide an omnibus (streaming?) upload endpoint. Or both.