Refine SRS timer and cond for performance issue.

[winlinvip]

Reference: ossrs/state-threads#13 (comment)

This is the current SRS UDP scheduling structure diagram. It is basically very efficient, so st does not appear in the bottleneck. The points that can be improved include:

1. The system only has one sleep, in the main coroutine, which is converted into a cond to notify other coroutines.
2. HTTP Stream uses sleep, so it will trigger ST switching performance. The performance of HTTP-FLV is lower than that of RTMP. It should use cond like RTMP.
3. MR (Merged Write) uses sleep, so the read optimization efficiency is not high. It should use readmsg or readmmsg, combined with cond processing. However, read optimization is different from write optimization. Write can determine how many messages to write after, while read is more troublesome, especially for TCP. It depends on how to implement port reuse, so that there can be many messages to read (such as RTMP back to the source using a private protocol).
4. Each Consumer has a cond, causing the Source to call the cond signal for each Consumer. In fact, the cond can be placed on the source, allowing each consumer to wait for the source's event, implementing broadcast, and batch awakening these coroutines.

[winlinvip]

I discovered that the ST (State Threads) underlying layer will also continuously call getsystime. Additionally, ST's epoll_wait(timeout=0) will have performance issues when there are many sleep calls. These issues can be optimized together.