Refactor: drop vestigial orch_built_on_host flag

[ChaoWao]

Summary

Runtime::orch_built_on_host_ used to distinguish the host-built-graph runtime
from the (since removed, #701) aicpu_build_graph one. With only
host_build_graph and tensormap_and_ringbuffer left it is a per-runtime
constant, and after #760 removed the last platform-layer caller it is also
unreferenced outside the runtime internals. This drops it:

• host_build_graph (a2a3 + a5): get_orch_built_on_host() was hard-coded
  to true and has no callers — deleted.
• tensormap_and_ringbuffer (a2a3 + a5): bind_prepared_to_runtime_impl
  unconditionally sets the flag to false before the runtime struct reaches
  any reader (the ctor's = true was always overwritten first), so every
  get_orch_built_on_host() site already took the device-orchestration path.
  Inlined that:
  • drop the orch_built_on_host_ field + getter + setter (runtime.h,
    runtime/shared/runtime.cpp) and the set_orch_built_on_host(false) call
    in host/runtime_maker.cpp;
  • aicpu/aicpu_executor.cpp: the if (get_orch_built_on_host()) { /* host orchestration, no-op */ } else { ... } becomes a plain scope (kept as a
    block so the per-callable dlopen / SO-table locals stay scoped); the
    SM-header spin-wait and the rt-destroy guard lose their always-true
    !get_orch_built_on_host() prefix;
  • runtime/scheduler/scheduler_cold_path.cpp: orchestrator_done_ init
    becomes a literal false.

No behavior change — every removed branch was statically unreachable.

Testing

• a2a3 + a5 sim libhost_runtime.so (both runtimes) rebuild cleanly;
  pre-commit (clang-format, clang-tidy, cpplint) passes.
• Hardware: a2a3 + a5 tensormap_and_ringbuffer ST (e.g. an AICPU-orch
  example) still runs to completion — the touched code is the AICPU
  executor's orchestrator/scheduler path.

[gemini-code-assist]

Code Review

This pull request removes the orch_built_on_host flag and its associated logic from the a2a3 and a5 runtimes, simplifying the AicpuExecutor and scheduler by consolidating the orchestration path. Review feedback suggests implementing RAII guards or try-catch blocks for resources like dlopen handles within the refactored orchestration scopes to prevent potential resource leaks.