Roland SP-404MK2: add support for reading and writing projects

[douglas-carmichael]

Adds a detector and creator for the Roland SP-404MK2, requested in #215.

The SP-404MK2 is a pad sampler, not a keyboard multi-sampler - a project holds 10 banks (A-J) of 16 pads, and each pad plays one sample; there are no key ranges. So, as with the other pad devices, each populated bank is converted to one multi-sample and each pad becomes a single-key zone.

Format

Not documented by Roland. It was reverse-engineered from real exported projects and the SP-404MKII v5.52 firmware image, and written up in documentation/design/SP404MK2_FORMAT.md:

• A project is a folder (PROJECT_XX) with a PADCONF.BIN pad configuration and a SMPL folder of BANK<bank>-<pad>.SMP samples.
• PADCONF.BIN (RFPD): 160 pad records of 172 bytes plus a name block. Both forms are read - the 31,488-byte export form and the 52,000-byte device-internal form (they differ only in the header size, which shifts the two blocks, and a trailing marks section). The export form is written, since that is what IMPORT PROJECT accepts.
• .SMP (RFWV): 32-byte header, 48 kHz / 16-bit, and the PCM is big-endian.
• Read and written per pad: the name, level, panning, loop on/off, pitch/fine tune.

Testing

Software-tested against two real exported projects (and a synthesized internal-form one):

• SP-404MK2 -> SP-404MK2 round-trips with byte-identical audio; the PADCONF.BIN header and all modeled record fields come back byte-identical to the device's own file, including the exact scaffold of an unused pad.
• Also tested: conversion from a foreign format (Bitwig) into the format and back, library/multi-bank mode, the mono processing option, bank overflow (a 170-zone source fills 160 pads and logs the rest as skipped), and malformed input (wrong size/magic, missing SMPL, corrupt .SMP, empty project) which is skipped without an exception.

Not yet verified on hardware - I do not own the device, so I matched the byte layout of real exports as closely as possible, but a real IMPORT PROJECT test is still outstanding. @allcounter offered hardware testing in #215.

Known limitations

• The on-device start/end trim and loop points are preserved in the file but not applied on reading - the full sample is always extracted. The field semantics are decoded (+0x04 play start, +0x2C loop start, byte offsets, untrimmed = 512), but the exact frame base could not be pinned down from the available projects, and extracting the wrong window would corrupt audio. Details are in the design doc.
• Patterns (PTN), pattern chains and project artwork are not converted.

Note

This PR is the format only. The small AbstractCreator.createSafeFilename change it originally carried has been split out into #226, since it affects the output of every creator and deserves to be judged separately. This PR does not depend on it - without #226 the SP-404MK2 simply shows an underscore instead of a dot in the few pad names that contain one (e.g. @1prod.mars Gala), and nothing else differs.