Handle DTS overlap and AV_NOPTS_VALUE inputs

[j0sh]

Summary

Fix DTS overlap between segments and handle AV_NOPTS_VALUE inputs that could
cause runaway encodes with millions of output frames.

Supersedes #423 and incorporates fixes for #443. Also addresses the root cause
behind the AV_NOPTS_VALUE-related issues discussed in #440.

Problem

Three interrelated timestamp problems in the transcoding pipeline:

1. DTS overlap between segments. When segments are processed out of order
   (e.g. flip-flopping between transcoders) or non-consecutively, drift
   accumulated from adjusting timestamps around the FPS filter could produce
   non-monotonic DTS across segment boundaries. This was unpredictable and
   depended on the segment processing order.

2. AV_NOPTS_VALUE causing runaway encodes. Certain inputs contain
   misplaced SEI NAL units (after picture data instead of before), which
   causes the ffmpeg H.264 parser to omit timing information, producing frames
   with AV_NOPTS_VALUE PTS. When a segment is skipped (e.g. segment 1 → 3),
   this could have led to PTS underflows in the FPS filter, generating
   millions of duplicate frames and filling disk.

3. Unclamped timestamps reaching the muxer and encoder. The decoder may
   still produce out-of-order PTS for pathological inputs. Without clamping,
   these regressive timestamps could reach the muxer or encoder, producing
   undefined behavior.

Changes

Filtergraph reset per segment (filter.c, encoder.c, transcoder.c)

Close and recreate the video filtergraph at the start of each segment instead
of persisting it across segments. This eliminates the drift from adjusting
input timestamps to satisfy the FPS filter's monotonicity requirement. The
performance cost is roughly 1% from filtergraph re-creation.

Because the filter resets each segment, a large amount of special-case
timestamp handling code for inter-segment discontinuities has been removed.
Some flush logic is retained for low-fps or very short content.

Decoded PTS repair (decoder.c)

Add fix_video_pts() at the decode stage to normalize video frame PTS before
it reaches the filter or encoder:

• Falls back to best_effort_timestamp when PTS is AV_NOPTS_VALUE
• Synthesizes PTS from frame duration + last PTS when both are missing
• Clamps regressive PTS to be strictly monotonic

Timestamp clamping at muxer and encoder (encoder.c)

• Clamp non-monotonic video DTS in the muxer independently of the existing
  DTS > PTS repair path, matching ffmpeg's own mux behavior
• Add a post-rescale tie-break guard for video PTS going into the encoder,
  handling cases where timebase conversion collapses adjacent frame PTS into
  the same encoder tick

SEI NAL reordering (sei_fixup.go)

Add a Go-side pre-processing step (FixMisplacedSEI) for H.264 mpegts inputs
that detects and reorders misplaced SEI NAL units (found after VCL NALs) to
appear before them. This restores the H.264 parser's ability to extract
timing information, eliminating the AV_NOPTS_VALUE frames at their source.

Runaway encode guard (encoder.c, transcoder.c)

Abort video encoding when output frames exceed 25x decoded frames
(lpms_ERR_ENC_RUNAWAY), providing a safety net against the FPS filter
exploding frame counts even after the other fixes. Image-sequence inputs
(image2) are exempted since frame expansion is expected there. (From #443)

Known Limitation

A DTS overlap can still occur between the first and second segments when all
of the following are true:

1. The first segment starts at or near PTS = 0
2. B-frames are in use (producing DTS < 0)
3. mpegts is the output format (cannot represent negative timestamps)

The mpegts muxer offsets all packets to compensate for negative DTS, but
subsequent segment transcodes are unaware of this offset. This is the same
behavior as FFmpeg CLI when muxdelay is set to 0, and is codified in
TestTranscoder_API_DTSOverlap. Adding a constant delay (as FFmpeg does with
its default 1.4s muxdelay) would fix this, but is deferred to avoid
accumulating delays across re-transcodes.

Test Plan

• TestTranscoder_API_DTSOverlap — verifies DTS monotonicity across
  out-of-order segments (and documents the known B-frame/mpegts edge case)
• TestTranscoder_NOPTS_SkipSegment — transcodes real-world samples with
  misplaced SEI and skipped segments that previously produced runaway output
• TestTranscoder_NOPTS_MissingSEIAndPES — verifies PTS synthesis for
  inputs where the H.264 parser cannot produce any timestamps
• TestTranscoder_EncodedFrameRunaway — triggers the 25x frame count abort
• TestFixMisplacedSEI_BrokenFiles / TestFixMisplacedSEI_NoChanges —
  unit tests for SEI reordering on affected and unaffected samples

[j0sh]

Ran a test with a go-livepeer B+O with around 12 hours' worth of customer segments being sent in alternating order, checked timestamps for all, and and things check out OK.

CI is green. Nvidia unit tests don't run in CI but pass locally.

Merging this so we can cut a release and get the fix into the wild.