`Reader.prototype.addChunk` was calling `Buffer.concat` constantly, which
increased garbage collection and just all-around killed performance. The
exact implications of this is documented in
brianc/node-postgres#1286, which has a test case
for showing how performance is affected.

Rather than concatenating buffers to the new buffer size constantly, this
change uses a growth strategy that doubles the size of the buffer each
time and tracks the functional length in a separate `chunkLength` variable.
This significantly reduces the amount of allocation and provides a 25x
performance in my test cases, the larger the amount of data the query
is returning, the greater improvement of performance.

Since this uses a doubling buffer, it was important to avoid growing
forever, so I also added a reclaimation strategy which reduces the size
of the buffer wever time more than half of the data has been read.

1. Fix Node @v0.10.x by removing `Buffer.alloc` usage.
2. Remove uneeded Reader.prototype._save/
3. Reset buffer at the end of the packet instead of shrinking dynamically.
4. Remove some semi-colons to keep style consistent.