[zfleeman]

Closes #139.

When an encode lands over the target, twopass_loop() re-encodes at a lower bitrate. Each retry used to run both passes from scratch in a new temp folder. The first pass only studies the video and writes stats, and a retry changes nothing but the bitrate, so it wrote the same stats again. Now a retry skips the first pass and runs the second pass with the stats from the first attempt.

This only affects software codecs. Hardware encoders (*_nvenc, *_videotoolbox) run a single pass already.

Changes

• twopass.py: TwoPass.run() takes two optional arguments. log_dir is the folder for the stats and progress files (without one, run() makes and deletes its own temp folder as before, e.g. for --dry-run). reuse_first_pass skips the first pass and prints "Reusing the first pass's stats". A pass's progress file is deleted before the pass starts, so a retry's progress bar doesn't briefly show the last attempt's progress.
• __main__.py: twopass_loop() creates one TemporaryDirectory for all attempts and passes reuse_first_pass=True from the second attempt on. EncodeJob.start_pass() counted a new attempt when a first or single pass started. A retry now starts with a second pass, so a second pass right after another second pass also counts as a new attempt. The retry's progress bar runs from 50% to 100%, since the first half is already done.
• README.md: a sentence in the codec notes about faster retries, with the numbers below.

Checked

10 seconds of a 1080p60 Call of Duty clip. Each codec encoded at 6 MiB, then retried at 5.1 MiB twice: once reusing the first attempt's stats, and once from scratch.

Codec	Retry with reused stats	Retry from scratch
x264	5.07 MiB in 4.0 s	5.13 MiB in 6.9 s
x265	4.85 MiB in 10.4 s	4.92 MiB in 23.0 s
vp9	5.12 MiB in 44.1 s	5.12 MiB in 60.3 s

The reused retries land about as close to the target as full re-encodes. With verbose logging, x265 logs slice=P but 2pass stats say B and a B-frame warning on every plain two-pass encode, with or without this change, and a reused retry adds no new warnings.

A CLI run at -s 5.1 with x264 went over on the first attempt (5.12 MB), reused the stats, landed at 5.02 MB, and left no ff4d- temp folders behind.

AV1 isn't tested: this machine's ffmpeg build doesn't include libaom-av1. libaom's first-pass stats don't depend on the target bitrate either, so it should behave like VP9.

Tests

• test_twopass.py: with reuse_first_pass, only the second pass runs and it points at the first attempt's stats (-passlogfile for x264 and VP9, stats= for x265). A caller's log_dir is used and not deleted. Hardware codecs are unchanged. A retry's progress file starts empty.
• test_main.py: twopass_loop() passes reuse_first_pass as False, True, True across three attempts with one shared stats folder, which is deleted afterward. EncodeJob.status() counts attempt 2 for a retry that starts with the second pass, and for hardware retries.