attached an updated patch cleanly applying to head, and raising cache capacity to 512 bytes.

(file #14608)

This patch implements a fixed-size cache in gnash::stream class
filled on first bitwise read and used till empty.
The patch gave me no regressions and dramatically
reduced tu_file based seeks (sometimes triggering complete
decompression from start).

With this patch loading the ninja movie performs 80k calls
to low-level input rather then 18m (1/1000). Only tested the
case in which low-level call is std_read_func (loading from filesystem), no idea what would that mean for network-loaded
SWFs.

Note that the bitwise reading functions in this patch are not
optimized so you don't see a big speed improvement, anyway
the patch provides a cache which may be used to implement those
improvements.

I'm suspending further work on it as I think we need an
to take a top-down design of data flow to really optimize
buffers (to avoid the data to be copied in memory between
buffers).