Xiaoya refactor block

[xiaoyachong]

The main changes of this PR

1. Use the arroyopy block YAML format to replace the processor_cli_tpx.py.

2. Define XPSResultStart, sent from XPSOperator to XPSWSResultPublisher, with the workflow:

ZMQ wire (splash_timepix / LabVIEW)
↓
(XPS operator) XPSTimepixZMQListener / XPSLabviewZMQListener
↓
(XPS operator) XPSStart / XPSRawEvent / XPSStop
↓
(XPS operator) XPSOperator
↓
(XPS operator) XPSResultStart / XPSResult / XPSResultStop
↓
(XPS operator) XPSWSResultPublisher
↓
(arroyo operator) XPSWebSocketListener
↓
LSE pipeline

3. Calculate running_mean for flush n as follows, and I check with Johannes that the calculation is correct in this splash_timpix issue:

* Flush 1: Array Data = sum of cycles 1–1000
  cycles_in_flush = 1000, total_cycles = 1000
  cumulative_sum = flush1
  running_mean 1 = cumulative_sum / 1000

* Flush 2: Array Data = sum of cycles 1001–2000
  cycles_in_flush = 1000, total_cycles = 2000
  cumulative_sum = flush1 + flush2
  running_mean 2 = cumulative_sum / 2000

* Flush 3: Array Data = sum of cycles 2001–3000
  cycles_in_flush = 1000, total_cycles = 3000
  cumulative_sum = flush1 + flush2 + flush3
  running_mean 3 = cumulative_sum / 3000

4. Test with splah_timepix simulator and LSE, and everything works well.

Issues with this PR

The splash_timepix simulator currently generates one flush per scan, which does not cause issues for LSE, since the model inference time is around 0.5 s per image on my MacBook. However, Johannes mentioned that in real experiments it is possible to generate hundreds or thousands of flushes. In that case, LSE may run into performance issues. We may need to add some accumulation logic in ArroyoXPS to limit the data frequency sent to LSE.

Companion PRs:
LSE: mlexchange/mlex_latent_explorer#76
arroyopy: als-computing/arroyopy#18

[taxe10]

Just a couple of very minor comments. Looks good.

[dylanmcreynolds]

I'm ok to merge this. Great work!