fix: Seen circuits may be accepted again

[lmondada]

It seems I have found a bug in multi-threaded TASO. Maybe this helps explain the bad performance?

I have run a test on 4 cores, 15seconds timeout, the results are marginally better:

on main

Compiling rewriter...
Using 4 threads
Optimising...
new best of size (major=192, minor=326)
new best of size (major=190, minor=324)
new best of size (major=188, minor=322)
new best of size (major=186, minor=320)
new best of size (major=184, minor=318)
new best of size (major=182, minor=316)
new best of size (major=180, minor=314)
Timeout
Optimisation finished
Tried 19447 circuits
END RESULT: (major=180, minor=314)
Joining worker threads
Saving result
Peak memory usage: 6.131306 GB
Done.

this PR

Compiling rewriter...
Using 4 threads
Optimising...
new best of size (major=192, minor=326)
new best of size (major=190, minor=324)
new best of size (major=188, minor=322)
new best of size (major=186, minor=320)
new best of size (major=184, minor=318)
new best of size (major=182, minor=316)
new best of size (major=180, minor=314)
new best of size (major=178, minor=312)
Timeout
Optimisation finished
Tried 18765 circuits
END RESULT: (major=178, minor=312)
Joining worker threads
Saving result
Peak memory usage: 7.384696 GB
Done.

[aborgna-q]

Nice catch!

The diff doesn't show it, so for reference: hashed_circs was being sent without dropping the seen circs after the loop.

[aborgna-q]

Move the progress after the circuit_cnt increase (otherwise this may cause duplicated logs if circ_cnt is a round number.

[aborgna-q]

Same here,

Suggested change

	logger.log_progress(circ_cnt, None, seen_hashes.len());
	circ_cnt += 1;
	circ_cnt += 1;
	logger.log_progress(circ_cnt, None, seen_hashes.len());

[lmondada]

Nvm, after merging in #156 this was fixed anyways.