Unroll sscanf into strncmp and strtol

[msg7086]

Most of the time in parse_index was spent in sscanf. I unrolled it into strncmp and strtol to improve performance.

(5s -> 1.3s for a 330MB lwi file.)

[Asd-g]

If you have only LWLibavVideoSource(), the index file is always recreated instead to be parsed.

[msg7086]

Let me investigate

[msg7086]

Fixed the corner case of PTS and DTS = LLONG_MIN

[msg7086]

I might as well try to abstract and pull the lwi file IO part out of the lwindex.c.

[Asd-g]

Tested with ~1h video+audio. The parsing time is improved from ~2.66s to ~1.61s.

[msg7086]

On my computer (Ryzen 7940H)

All tests passed!

Performance Benchmark (2000000 iterations):
----------------------------------------   
Original main_index:    1932367.15 ops/sec
Unrolled main_index:    39215686.27 ops/sec (20.3x faster)
Original video_index:   2096436.06 ops/sec
Unrolled video_index:   48780487.80 ops/sec (23.3x faster)
Original audio_index:   5681818.18 ops/sec
Unrolled audio_index:   250000000.00 ops/sec (44.0x faster)
----------------------------------------

On ARM64 VPS (Neoverse-N1)

All tests passed!

Performance Benchmark (2000000 iterations):
----------------------------------------
Original main_index:    2469129.71 ops/sec
Unrolled main_index:    23321439.40 ops/sec (9.4x faster)
Original video_index:   2811749.18 ops/sec
Unrolled video_index:   31693209.73 ops/sec (11.3x faster)
Original audio_index:   8242052.60 ops/sec
Unrolled audio_index:   140567894.29 ops/sec (17.1x faster)
----------------------------------------

Test & benchmark code https://gist.github.com/msg7086/f0cf87f73b4e4affa5e02a4f7c2973f4