lcov: Reduce load times for complex coverage data files

lcov uses a significant amount of time loading coverage data files
containing complex branch coverage data (i.e. data with a large number
of branches per basic block). Most of this time is spent storing
branch coverage data in a vector-based data representation, with an
unnecessary amount of cross-checking being done for existing branch
data.

Fix this by replacing the vector based data representation by two
separate representations, scalar for storage and hash for processing,
and by moving cross-checking out of the hot path. This results in a
significant speedup at the cost of a minor increase in memory usage.

Test results for "make test SIZE=large":

Original:
  17 tests executed, 17 passed, 0 failed, 0 skipped (time 1883.9s, mem
  2459.0MB)

Patched:
  17 tests executed, 17 passed, 0 failed, 0 skipped (time 283.6s, mem
  2544.2MB)

Note that this fix only applies to the lcov tool. The same work is
necessary for genhtml.

This approach was inspired by a patch by creich.3141592@gmail.com.

Signed-off-by: Peter Oberparleiter <oberpar@linux.vnet.ibm.com>

Author: oberpar