fix(rdb_load): fix partial reads dropping elements

[andydunstall]

Found a bug in loading huge objects where we append to the wrong encoding.

Say you have a hmap with 4100 integers, we'll load 4092 then 8 elements. The first CreateHMap will see 4092 > 64 so create a StringMap, but the second CreateHMap will see 8 < 64 so creates an lpNew instead of appending to the existing StringMap (meaning the existing elements are lost). Same applies to set (when loading an 'intset') and zset.

Therefore adding an item.streamed to be really explicit about whether the item is being streamed so the larger encoding must always be used. We could probably infer from append and reserve but seemed better to be explicit.