Limit Gcs.Write to single storage.create api call

[relud]

to remove any chance of partial write complications and minimize the number of api calls.

requires #1237 to prevent data loss by ensuring that content.write is never called after content.toByteArray()

[jklukas]

Am I correct with context here that you've determined the GCS SDK is not releasing memory as we make .write calls, so there's no memory benefit to incrementally pushing data into the SDK, so we might as well do the simpler thing of accumulating a byte stream and flushing it to GCS all at once?

[relud]

Am I correct with context here that you've determined the GCS SDK is not releasing memory as we make .write calls, so there's no memory benefit to incrementally pushing data into the SDK, so we might as well do the simpler thing of accumulating a byte stream and flushing it to GCS all at once?

yes, but the reason is that with BATCH_MAX_DELAY=10s the sum of in-flight batches is at most around 30MiB uncompressed (not counting any copies lingering from concatenating byte strings).