Allocate before loading when responding to queries

In the [query executor](https://github.com/ipfs/go-graphsync/blob/c96b3b2944cacb405594dc15f2333145b2db133a/responsemanager/queryexecutor.go#L87), we seem to _load_ first, then allocate when sending the block. Unfortunately, when processing many parallel queries, we can end up loading a lot of blocks then _blocking_ (waiting to be able to "reserve space" for them. we'd likely use less memory if we just sent the block.

Instead, we should allocate first. That is:

1. Before loading anything, allocate 2-4MiB (maximum block size).
2. Then load.
3. Then reduce the allocation to the _actual_ amount of memory needed.

This means we:

1. Won't go over the allocation limit.
2. If we're blocked on allocating, we won't sit there while holding on to buffers.