Refactor: drop unused READY/RUNNING task states

[poursoul]

Summary

The PTO2TaskState enum advertised the chain
PENDING → READY → RUNNING → COMPLETED → CONSUMED, but in practice:

• PTO2_TASK_RUNNING was never written or read anywhere in src/.
• PTO2_TASK_READY was written by a single CAS in the profiling-instrumented
  release_fanin_and_check_ready and had no readers. The fanin_refcount.fetch_add already elects a unique winner for the ready-queue push, so the CAS was redundant; the non-profiling overload had already been doing this without writing task_state.

The actual transitions performed by the runtime are
PENDING → COMPLETED → CONSUMED; ready-vs-running is derived from
fanin_refcount / per-core running_slot_state, not from task_state.

Changes

• Drop PTO2_TASK_READY and PTO2_TASK_RUNNING; renumber the enum to
  PENDING(0) / COMPLETED(1) / CONSUMED(2). The < / >= PTO2_TASK_COMPLETED
  ordering comparisons in scheduler/orchestrator still hold under the new
  numbering since CONSUMED remains the largest value.
• Drop the redundant CAS PENDING → READY in the profiling-instrumented
  release_fanin_and_check_ready so it matches the non-profiling path.
• Delete the unused task_state_name() helper (no callers in repo).
• Update scheduler_cold_path stall-dump comment to reflect that task_state
  has no intermediate ready/running value.
• Update unit tests to use PENDING in place of READY/RUNNING fixtures
  (semantically equivalent: anything < COMPLETED).
• Sync state-machine docs: docs/orchestrator.md, docs/scheduler.md, and
  src/{a2a3,a5}/runtime/tensormap_and_ringbuffer/docs/RUNTIME_LOGIC.md.

Applied to both a2a3 and a5 runtime copies.

Test plan

• cmake --build build/ut_cpp --target test_scheduler_state test_task_state test_wiring
• ctest -R 'scheduler_state|task_state|wiring' — 3/3 pass
• rg 'PTO2_TASK_(READY|RUNNING)|task_state_name' — zero hits across src/, tests/, docs/

[gemini-code-assist]

Code Review

This pull request simplifies the task state machine by removing the explicit READY and RUNNING states from the PTO2TaskState enumeration. Tasks now remain in the PENDING state from submission through dispatch until they reach the COMPLETED state. Readiness is now determined by fanin_refcount, and the running status is tracked via per-core pointers. The changes include updates to documentation, the scheduler logic, and unit tests to align with this streamlined lifecycle. I have no feedback to provide.