cachegrind: ignore leading path to the benchmarked rustls checkout

[aochagavia]

For each PR benchmark we create two checkouts of the rustls repository, each one in its own directory (e.g. /tmp/.tmpZE2KDR and /tmp/.tmpq8vPPQ). As a consequence, when comparing the cachegrind output the diff tool fails to correlate the functions of both runs. The result is a totally useless diff, where all instructions from the baseline are considered to have disappeared and all instructions from the candidate to have appeared (for an example, see https://bench.rustls.dev/comparisons/371463d8129f4f0541e38063358ec2ad4428a44a:453cf7968b251b22db5bcadebefe3ea37125b630/cachegrind-diff/handshake_tickets_aws_lc_rs_1.2_rsa_aes_server)

This commit tells cachegrind to ignore the leading path (everything up to target/release/build), restoring the ability to track individual rustls functions.

[aochagavia]

@cpu I can't think of an easy way for you to test this... The way I developed the solution myself was by:

1. Manually running the icount benchmarks locally from two different paths (rustls and rustls2).
2. Creating a comparison with the helper code from rustls/ci-bench and manually checking the diffs in the output (cargo run --release -- compare foo bar). This reproduced the issue.
3. Iterating on a fix in the helper code.
4. Creating a comparison with the helper code and manually checking the diffs in the output (cargo run --release -- compare foo bar). This proved that the issue was solved.
5. Copying the fix to this repository.

[cpu]

This fix makes sense to me. I trust your debugging process here and don't think I need to reproduce it from scratch. If there are further issues I know where to look to iterate on the regex :-)

Thanks!