Sync/Wait for processing handler (#14)

Use a key-specific wait for delayed additions into the handler map.

fix receive race (#13)

Move close and recycle of req into waiter's code

Fix send/receive synchronization race condition (#12)

* Fix send/receive synchronization race condition

* restore cleanup and submitJob name

* Rename submitJob

* Add locking of job submission to single-thread it

* Add some logging

* Fix premature dequeuing issue

* Remove extra logging

* Remove submit locks

* Remove drain debug logging

* rename constants and change queue size to 8

Fix client tests (#8)

* Fix client tests

* Reverted to use of NewConnected
* Fixed arguments on some New Client calls

* Optimize conn checking in readLoop

* Add error handling for gearmanClient.Do() call

switch ssmccoy to quantcast.

client: Remove unnecessary dereferencing

Removed the unnecessary dereferencing of the pointer to the connection
struct.

client: major overhaul of buffer management

This still maintains a double-copy which is less than ideal, but is
significantly faster as a result because the target region which holds
the `data` parameter to the response is now dynamically allocated so as
to avoid unnecessary copies and dynamic/exponential allocation.

Further cleaned-up the parsing of the response and removed the
`decodeResponse` function which had lots of funky logic around wierd
buffer management.

Abstracted the reconnect behavior out into a separate method so it can
be reused whether the header or buffer are being read.

client: gearman client now faster and better

Eliminated the strange usage of in favor of using channels. Removed the
redundant `processLoop` goroutine, the `readLoop` now inlines the
processing (it was serial anyway!). Switched to the use of threadsafe
memory object pools for request and response objects, and now allows
pipelining if multiple goroutines are making requests at once.

Removed a lot of redundancy in the code, and eliminated the use of the
strange `responseHandlerMap` in favor of a concurrent map for holding
callbacks on a per-handler basis.

Refactored the `request` type such that it now uses a builder pattern,
and now itself keeps a small byte buffer for each request which is
reused when parameters are attached to requests. Streamlined the
formatting of outbound messages such that they are encoded as they are
written to the output buffer, to reduce the number of copies involved in
most operations.

Eliminated the strange blind-read loop which pretended it had no idea
how large of a buffer it would need to allocate to read a response, and
now allocates a buffer for each payload, the size of the payload as
specified by the server, respectively.

client: gearman client now faster and better

Eliminated the strange usage of in favor of using channels. Removed the
redundant `processLoop` goroutine, the `readLoop` now inlines the
processing (it was serial anyway!). Switched to the use of threadsafe
memory object pools for request and response objects, and now allows
pipelining if multiple goroutines are making requests at once.

Removed a lot of redundancy in the code, and eliminated the use of the
strange `responseHandlerMap` in favor of a concurrent map for holding
callbacks on a per-handler basis.

Refactored the `request` type such that it now uses a builder pattern,
and now itself keeps a small byte buffer for each request which is
reused when parameters are attached to requests. Streamlined the
formatting of outbound messages such that they are encoded as they are
written to the output buffer, to reduce the number of copies involved in
most operations.

Eliminated the strange blind-read loop which pretended it had no idea
how large of a buffer it would need to allocate to read a response, and
now allocates a buffer for each payload, the size of the payload as
specified by the server, respectively.

Release 0.5.1

- Fixes robfig/cron bug that causes multiple invocation of scheduled function.
- Build binary inside Alpine Docker image.