perf(coverage): Improve coverage collection by using merge

[nathaniel-brough]

This approach creates a dummy-corpus and merges the main corpus into the dummy-corpus. After the corpus is merged we delete the dummy-corpus. This allows us to calculate code coverage signficantly faster than calculating it per-file in the corpus, as it can be run in parallel and there is less startup initialisation cost for creating a new process for each entry in the corpus.

This was mainly inspired by the approach taken in google/oss-fuzz.

This should also fix #254 as there will only be one set of profile data per corpus directory rather than a set of profile data per file in all the corpus directories. Although I can't confirm this as I don't own a mac :)

Performance diff:

Testing performance on the fuzzers in quick-xml. With a corpus of 38780 files.

Before

time cargo fuzz coverage
# ...
real    43m51.644s
user    9m10.208s
sys     11m41.468s

After

time cargo fuzz coverage
# ...
real    0m25.826s
user    1m14.378s
sys     0m6.698s

Delta

real: 101x faster

[nathaniel-brough]

I've also tested this change to ensure that the same coverage is calculated before and after the change.

 $ rm fuzz/coverage/fuzz_target_1/coverage.profdata
 $ cargo fuzz coverage -O fuzz_target_1
 $ cargo cov -- report fuzz/target/x86_64-unknown-linux-gnu/release/fuzz_target_1 -instr-profile fuzz/coverage/fuzz_target_1/coverage.profdata
 src/