Add mrb_task_set_scheduler_hook for pre-scheduling deferred work. Remove mrb_hal_task_switch_hook instead - #6982

## Motivation

PicoRuby is building an ISR-to-task event bridge: interrupt handlers
only stage work (set flags), and a deferred handler wakes tasks
blocked in Task::Queue#pop. This needs a servicing point that runs in
thread context BEFORE the scheduler reads the ready queue, so a task
woken there is picked up in the same iteration. Today mruby/c has no
such point: mrbc_hal_idle_cpu() runs only when the ready queue is
empty, and nothing runs before task selection.

This mirrors mruby's mrb_task_set_scheduler_hook (mruby/mruby#6982);
PicoRuby runs both VMs and keeps their scheduler surfaces symmetric.

## How it works (UART RX as the example)

The ISR does two things only, neither of which touches the VM:

    void uart_rx_isr(void)
    {
      ring_buffer_put(uart_getc());  /* data into a plain C buffer */
      rx_flag = 1;                   /* volatile flag: "something arrived" */
    }

The hook is the deferred second half:

    void my_hook(void *ud)
    {
      if (rx_flag) {
        rx_flag = 0;
        mrbc_task_queue_push(&q, &ev);  /* wakes the task blocked in q.pop */
      }
    }

    /* registered once at init */
    mrbc_task_set_scheduler_hook(my_hook, NULL);

Timeline for the important case -- every task asleep:

    task B:    ev = q.pop .......... queue empty, parks as WAITING
    scheduler: no READY task -> mrbc_hal_idle_cpu()
                    |
       UART interrupt fires!
       ISR: byte into buffer, rx_flag = 1, returns  (VM untouched)
                    |
    scheduler: returns from idle, reaches the loop top
       hook: sees rx_flag -> queue push -> task B becomes READY
       scheduler: reads the ready queue right after -> picks B
    task B:    returns from q.pop, drains the ring buffer

An interrupt can fire at any moment -- possibly mid-allocation -- so
an ISR must never push into a Task::Queue. The hook runs at a moment
where the VM is by construction quiescent (between tasks, right
before the ready-queue read), so it carries the ISR's note across to
the safe side: the ISR leaves a mark, the hook performs the regular
push.

## Why pushing from the hook is safe

- Thread context: mrbc_task_queue_push forbids interrupt context and
  VM re-entry (see the comment in c_task_queue.c); the hook's call
  sites are the bodies of mrbc_run/mrbc_run_step, which are neither.
- No allocation is in flight at the loop top, so the push's array
  growth cannot corrupt an interrupted allocation.
- The TCB relinking inside the push runs under the same
  mrbc_hal_disable_irq/enable_irq exclusion as any thread-context
  caller; the only concurrent mutator is the tick IRQ, which that
  exclusion covers.
- No flag_preemption is needed: the very next statement reads the
  ready queue, so a task woken here is selected immediately.

## What this adds

- mrbc_task_set_scheduler_hook(fn, ud): a process-global hook invoked
  in thread context right before the ready-queue read, at both
  scheduler entries (the mrbc_run loop top and mrbc_run_step).
  mruby's hook is per-mrb_state; mruby/c's is process-global because
  the scheduler queues are process-global (VMs live per TCB).
- The whole feature sits behind #if defined(MRBC_TASK_SCHEDULER_HOOK):
  default builds compile it out entirely -- zero cost and zero
  behavior change for existing applications and HALs. Enabled builds
  pay one pointer check per scheduler iteration while no hook is set.
- Single-owner contract (documented in rrt0.h): setting replaces,
  fn == NULL clears, composition is the embedder's concern, the
  caller owns ud. The hook must be cheap, never sleep, and never
  re-enter the scheduler.
- No HAL involvement: gems/embedders register at init, so no port
  needs changes.

Behavioral tests for the contract live in PicoRuby's integration
suite; the mruby side carries equivalent tests in mruby-task's own
mrbtest suite, while mruby/c's pure-Ruby test harness cannot register
a C hook.

mruby/mruby#6982 replaced mrb_hal_task_switch_hook with
mrb_task_set_scheduler_hook; the submodule already points at the
merged master. Follow up on the PicoRuby side:

- picoruby-machine/ports/rp2040/machine.c: the cyw43_arch POLL pump
  moves from the removed switch hook into a scheduler hook registered
  in picorb_hal_init (PICORB_VM_MRUBY + PICO_CYW43_ARCH_POLL only).
  Same cyw43_is_initialized() guard, same cheap/no-sleep contract.
  The scheduler hook fires at every scheduler entry, so the busy-path
  pumping cadence is preserved and idle iterations now pump as well.
- picoruby-mruby/ports/rp2040/task.c, ports/esp32/task.c: removed;
  they only implemented the deleted HAL symbol.

Verified builds: rake picoruby:debug and femtoruby:debug (host),
r2p2:picoruby:pico2_w:debug (mruby + PICO_CYW43_ARCH_POLL -- the
affected configuration), and r2p2:picoruby:pico2:debug (no cyw43).
On-device revalidation of the #6942 busy-loop/GC-drain scenarios is
still pending.