applet.benchmark: skip going through the MODE state. This improves the throughput a bit. (before) I: glasgow.applet.benchmark: mode source: 16.999 MiB/s I: glasgow.applet.benchmark: mode sink: 18.138 MiB/s (after) I: glasgow.applet.benchmark: mode source: 23.196 MiB/s I: glasgow.applet.benchmark: mode sink: 23.840 MiB/s

firmware: update libfx2.

access.direct.demultiplexer: auto-hint reads with length.

support.logging: implement dump_hex, and use in USB and FIFO code. Most of the time it is not necessary (and in fact is harmful) to eagerly dump the complete packet at USB and FIFO layers. Instead, lazily dump only some first bytes, which is sufficient for most debugging. In any case, `glasgow -vvv` dumps the complete data.

support.lazy: implement.

applet.benchmark: hint reads to use total length. When making very large reads, it is often possible to either predict the total length exactly (e.g. the shugart-floppy applet) or to use a "sufficiently large" buffer (e.g. the lpc applet). Thus, it seems that a benchmark should not use an artificially small hint.

firmware: add Device Qualifier request. Necessary for USB compliance.

access.direct.demultiplexer: use ChunkedFIFO. This significantly improves `benchmark sink`: (before) I: glasgow.applet.benchmark: mode sink: 1.454 MiB/s (after) I: glasgow.applet.benchmark: mode sink: 16.420 MiB/s On my system the sink rate asymptotically approaches around 20 MiB/s. 16 MiB/s is an approx 10x improvement, and uses 1/32 of the Linux usbdevfs memory limit, which is acceptable for now. Fixes #79.

support.chunked_fifo: fix bool() with partial reads.

support.chunked_fifo: implement. Required for #79.