Change fastcdc to a better and simpler algorithm. (#79)

This CL changes the chunking algorithm from "normalized chunking" to
simple "regression chunking", and changes the has criteria from
'hash&mask' to 'hash<=threshold'. These are all ideas taken from
testing and analysis done at
  https://github.com/dbaarda/rollsum-chunking/blob/master/RESULTS.rst
Regression chunking was introduced in
  https://www.usenix.org/system/files/conference/atc12/atc12-final293.pdf

The algorithm uses an arbitrary number of regressions using power-of-2
regression target lengths. This means we can use a simple bitmask for
the regression hash criteria.

Regression chunking yields high deduplication rates even for lower max
chunk sizes, so that the cdc_stream max chunk can be reduced to 512K
from 1024K. This fixes potential latency spikes from large chunks.

manifest/manifest_updater.h

@@ -51,7 +51,7 @@ struct UpdaterConfig {
   size_t avg_chunk_size = 256 << 10;
 
   // Maximum allowed chunk size.
-  size_t max_chunk_size = 1024 << 10;
+  size_t max_chunk_size = 512 << 10;
 
   // Size of the chunker thread pool. Defaults to the number of available CPUs.
   uint32_t num_threads = 0;