Previously h2load used default flow control window as described in
HTTP/2 and SPDY specification.  The window size is 64KiB, which is a
bit small, and cannot utilize full server performance when response
size is not too small.  Basically, we do this kind of benchmarking
test to measure server's throughput, and optimal performance.  Smaller
window certainly degrades performance even in local testing because
server is so fast that it has to wait for WINDOW_UPDATE from h2load.
To make default behaviour suitable for peak performance test, we
decided to disable flow control in h2load by setting large enough
window size.

Most users used h2load without -w or -W options, so they were
implicitly throttled by flow control and the result was affected by
that negatively.  Now flow control is disabled by default, the result
may improve depending on the implementations.

After reviewing codebase, only queue for DATA frames requires
priorities.  Other frames can be replaced multiple linear queues.
Replacing priority queue with linear queue allows us to simplify
codebase a bit; for example, now nghttp2_session.next_seq is gone.

Since application most likely allocates the stream object in
nghttp2_on_begin_headers_callback, it is desirable to handle its
failure as stream error.  But previously it only signals success or
fatal error.  Submitting RST_STREAM does not prevent
nghttp2_on_header_callback from being invoked.  This commit improves
this situation by allowing NGHTTP2_ERR_TEMPORAL_CALLBACK_FAILURE from
nghttp2_on_begin_headers_callback.  If that value is returned, library
submits RST_STREAM with error code INTERNAL_ERROR, and
nghttp2_on_header_callback and nghttp2_on_frame_recv_callback for that
frame are not invoked.  Note that for PUSH_PROMISE frame, the stream
to be reset is promised stream.

Implement fd caching for static files.  The response body for such as
404 was dynamically generated previously, but now it is written in
temporally file and its fd is cached.  Currently, cache is reference
counted and expired when count becomes 0.  This makes caching is not
effective other than "busy" period, but we don't need this feature if
we are not busy.

In this commit, we made --dep-idle behaviour by default.  This is
because the previous default behaviour is not reflect current usage of
dependency priority and never will be because of fragility of tree due
to stream closure.