[Performance] Runtime performance optimization tracking

[zhusy54]

This issue tracks runtime performance optimization work for the tensormap_and_ringbuffer runtime on a2a3 / a5. Findings, measurements, and proposed optimizations are added as comments.

Orchestrator-specific tracking is split out to #984 — covers a2a3 per-phase baseline measurements, a5 platform pressures (cluster topology, frequency, AICore +50%), Host-on-UB deployment evaluation, and dedicated-hardware design sketches for the streaming orchestrator. This issue (#545) continues to track the rest of the runtime path: scheduler / dispatch, completion handling, runtime wrapping (Python / C bindings), inter-task synchronization, AICPU thread placement, etc.

Landed optimizations

• Dual-slot scheduling for mix subgraph tasks ✅ — PRs Add: dual-slot AICPU dispatch payload and two-phase pipelining scheduler #477, Refactor: unify two-phase dispatch and fix sync_start spin loop #553. AIC + AIV kernels inside a mixed task now dispatch into two hardware slots concurrently instead of serializing.

Reproduction

python examples/scripts/run_example.py \
    -k tests/st/a2a3/tensormap_and_ringbuffer/paged_attention/kernels \
    -g tests/st/a2a3/tensormap_and_ringbuffer/paged_attention/golden.py \
    -p a2a3 -d 5 -n 10

[ChaoWao]

For A5, we can try

1. 1orch+3sch的性能（current）
2. 同cluster内部两个cpu，1orch+1sch的性能
3. 使能1个cluster内的两个cpu的smt，这样变成4个cpu线程，测1orch+3sch的性能
4. 使能1个cluster内的一个cpu的smt，这样变成3个cpu线程，测1orch+2sch的性能

[ChaoWao]

SPMD task dispatch has 18u delay between the first and the last task.

[ndleslx]

For worker.run, the host time = device + python/c calling. The python/c calling accounts for about 3ms which should be hidden by device run.