worker: avoid reading 1 byte at a time from the OS pipe Apparently `pickle.load` does a lot of small reads, many of them literally 1-byte, so it benefits greatly from buffering. This change enables the buffering, at the cost of more complicated interaction with the `selector` API. On one repository with ~400k files this reduces the time by about ~30s, from ~60 to ~30s. The difference is so large because the actual updating work is parallellized, while these small reads are bottlenecking the central hg process.