[h-mayorquin]

ABF files do not store the command that the amplifier played as samples. They store a protocol: a holding level plus a table of up to ten epochs per output, each with a shape, a level, a duration and how these change from sweep to sweep. read_raw_protocol rebuilds the command from that table, and right now it has these problems:

• Every epoch is drawn as a step, so ramps and the pulse, triangle, cosine and biphasic trains come out as flat levels.
• ABF1 files throw an error, although their header carries the same protocol.
• The "use last epoch level" option is ignored, so a sweep that should start where the previous one ended starts at the holding level.
• Files with user lists, stimulus files or alternating outputs get a wrong waveform without any message.

This PR draws all six epoch shapes with their train period and width, keeps the level and duration increments, follows the "use last epoch level" option and reads the protocol of ABF1 files from version 1.6 on. Files with user lists, stimulus files, alternating outputs or conditioning trains are not reconstructed yet, and the reader now warns about them. Before the first epoch the output stays at the holding level for 1/64 of the sweep. The old code used that value with a # TODO guess for first holding comment and #1305 asked about it. It is right (the recorded responses start right after it), so it stays. The test files were added to gin in https://gin.g-node.org/NeuralEnsemble/ephy_testing_data/pulls/197.

This should close #1305.