Add some sensible retry for NN uploads

[thomas-riccardi]

We previously explicitly choose to not do retry on error for NN uploads. Errors do happen and it's a shame to lose a full training for a temporary upload error.

However, we should have specific settings for that large upload:

• longer global timeout
• don't retry 10 times, maybe 3 is enough
• less retry, so more wait between retries (maybe 10s after first failure, then 30s?)

[thomas-riccardi]

See https://github.com/Deepomatic/vulcan/issues/1589