jefe: move a biggish bit of Jefe off the stack

[hawkw]

Currently, Jefe represents task states in an array that's initialized in the main() function and passed as a &mut [TaskStatus; NUM_TASKS] slice to the ServerImpl initializer. This is a kind of biggish array, since each TaskStatus is in excess of 64 bytes (it's a timestamp plus a couple enum discriminants), multiplied by NUM_TASKS, which can be in the vicinity of 20-30 tasks. Because the array is initialized as a local variable in the main() function, the array is on the stack all of the time. It doesn't have to be on the stack, and it could also be not on the stack.

This commit makes it not be on the stack by putting it in a static instead. Also, I made the static be initialized by code that runs in const fn, rather than having to loop over the held tasks at runtime and update them to be held by mutating the array. This just seemed a lot nicer, despite making the code look a bit less nicer. Oh well.

This makes Jefe's stack usage a lot smaller; contrast the Gimletlet image on master:

jefe: 2048 bytes (limit is 2200)
     [+0] _start
   [+840] main
   [+184] task_jefe::dump::dump_task
   [+776] task_jefe::dump::dump_task_run
    [+56] humpty::DumpAreaHeader::read_and_check
   [+136] task_jefe::dump::dump_task_run::{{closure}}
    [+40] <static_cell::StaticCell<ringbuf::Ringbuf<T,u16,_>> as ringbuf::RecordEntry<T>>::record_entry
    [+16] <humpty::DumpError<T> as core::cmp::PartialEq>::eq

...and after this change:

jefe: 1424 bytes (limit is 2200)
     [+0] _start
   [+216] main
   [+184] task_jefe::dump::dump_task
   [+776] task_jefe::dump::dump_task_run
    [+56] humpty::DumpAreaHeader::read_and_check
   [+136] task_jefe::dump::dump_task_run::{{closure}}
    [+40] <static_cell::StaticCell<ringbuf::Ringbuf<T,u16,_>> as ringbuf::RecordEntry<T>>::record_entry
    [+16] <humpty::DumpError<T> as core::cmp::PartialEq>::eq

So that's some 600B of main gone away, which seems cool.

We could probably go through and adjust all the Jefe task sizes to be smaller now that they can be smaller.

[hawkw]

oh, ugh. one major issue with this change is that this use of ClaimOnceCell breaks the build with the no-panic userlib feature, since the claim() method panics if the cell is already claimed (which never happens, but can't be statically eliminated). that's sad. maybe this change is a non-starter for now.

[hawkw]

(this is why we can't currently build this branch for donglet and any other app that builds jefe with no-panic enabled: https://github.com/oxidecomputer/hubris/actions/runs/20967849704/job/60263104877?pr=2351)