…ead_raw_eyelink_fix

pd.replace is very slow. Using numpy is 10x faster on a large file.
for example on a file with 4 million lines, pd.replace takes 30 seconds
and using numpy where takes 2.5 seconds.

Also nested the system message check in parse_recordings blocks under
the block of code where we know those messages appear. this way we dont
check for system messages on literally every line of the file!

new test creates a new temp file and adds additional channels so that more of
read_raw_eyelink will be touched in testing

fixed conflict in test_eyelink from 11826

Co-authored-by: Eric Larson <larson.eric.d@gmail.com>

now that we use the tmp_path fixture, no unlinking needed

Required update of tests to catch warnings