Stop writing trailing zeros out to chunk store

[bboreham]

We create all chunks in memory as 1K long, and write the full 1K to store every time, even though many chunks will be half-full or less so we are writing a lot of trailing zero bytes.

AWS then base64-encodes all those zeros.

There doesn't seem to be an API on Prometheus (v1) chunks to find out how many bytes are used, but we could simply look for the last non-zero byte in the Marshal() output. We already include the data length in Chunk.Encode().

It's not clear to me whether we need to add back all the zeros when we read back the data, for double-delta encoding.

Slightly related: #300

[bboreham]

Update: the compression code actually stashes some intermediate data at the end of the chunk, so not quite as simple as I thought. First reaction is to move to Prom2 chunks before fixing this, as they don't do that.

However, note that DynamoDB writes are costed in 1K increments, so to the extent we can get the whole write down below 1K it would be a win.