0.6.0 — the board's openkal implementation adopts 0.9.0, and is now exercised

[Sunrisepeak]

The pin was openkal 0.5.2, four minors behind. Adopting 0.9.0 changes three
things in src/kal/kal.cpp:

• Transfers are one signed word. kal_stream_write and kal_stream_read
  returned kal_io_result, a pair of a count and a condition, which 0.9
  withdrew: the pair was two machine words where one carries the whole answer,
  and every caller in the ecosystem was collapsing it by hand at the call site.
  A partial write that then fails now reports the COUNT and not the condition —
  the bytes reached the device, and a caller told kal_err_io would send them
  twice.
• kal_memory_granularity is new and answers 1. This machine has no memory
  management unit and this implementation imposes no rounding of its own, so
  every address and every length is acceptable, which the specification spells
  that way. ⚠️ It is not a page size, and the comment says so: openkal-musl
  0.7.0 took a granularity of one from another such implementation and assigned
  it to libc.page_size, whose allocator then asked for one-byte extents.
• kal_version and kal_interfaces are exported. They belong to no interface
  and every conforming implementation has them, so that a consumer with no
  linker to ask can ask before it calls. This one claims exactly abort, stream
  and memory.

And the part that was not a bump

⚠️ NOTHING CALLED ANY OF IT. tests/console.cpp calls board::* and nothing
else. The feature compiles src/kal/** and the linker takes all fourteen names
into the test binary, so nm shows them and the suite is green — which proves
the implementation compiles and links and says nothing about what it answers. A
wrong result in any of the three changes above would have linked and shown green
just the same.

tests/openkal.cpp asks. Nine observations, every one of something 0.9
introduced or changed, on both ISA profiles. Measured that they bite: with the
granularity returning 4096 and kal_interfaces claiming an interface this board
does not export, two of them fail and name themselves.

⚠️ It does not go quiet when the feature is off — it prints that it did not
observe anything and returns 0, and the workflow greps for the line that only
appears when it DID. A case whose "not run" and whose "passed" read the same is
not a case. Both widths assert it.