[QDP] 1211 biweekly community sync

[rich7420]

• Unit test and integration test
• latency benchmark
• I/O barrier
• python interface
• IPC pipeline

[rich7420]

Toy dataset benchmark

[400Ping]

set float32 to default

[400Ping]

Using allocation and double-buffered async I/O to solve I/O barrier issue

[rich7420]

I'd like to improve e2e benchmark through cleaning cache and memory between stages.

[rich7420]

I think we could merge norm and encode into a kernel method

[400Ping]

dynamically allocate cpu cores