Fix MVC per-view DPB management: add same_views mask

[intrepidsilence]

Summary

Two DPB management checks in edge264_headers.c count pending frames from both MVC views against per-view SPS limits. This causes incorrect frame bumping and output ordering in MVC streams.

The Two Bugs (Same Root Cause)

1. DPB fullness check (line ~1225)

// Before: counts both views against per-view limit
while (popcount(reference_frames | to_get & ~output) > max_dec_frame_buffering && max_bump--)

// After: counts only current view
while (popcount((reference_frames | to_get & ~output) & same_views) > max_dec_frame_buffering && max_bump--)

With both views contributing ~8 frames against a per-view limit of 4, the loop bumps too many frames, driving max_bump negative. The max_bump < 0 path then directly inserts frames into the FIFO output queue, bypassing bump_frame()'s POC-ordered output. This corrupts the base view's display order.

2. Reorder threshold check (line ~1231)

// Before: counts both views against per-view limit
} else if (popcount(to_get & ~output) > max_num_reorder_frames) {

// After: counts only current view
} else if (popcount(to_get & ~output & same_views) > max_num_reorder_frames) {

Counting both views' pending frames triggers premature bumping from the current view's queue. Since edge264_get_frame() pairs base and dependent views by FIFO queue position, this desynchronizes the two queues, pairing frames from different temporal moments.

Symptoms

• Bug 1: Base view (left eye) frame stutter during playback — frames appear out of display order
• Bug 2: Dependent view (right eye) exhibits ±1 frame temporal oscillation in a repeating 6-frame cycle matching the IBBBP GOP substructure

Quantitative analysis of a commercial MVC stream showed ~63% of dependent view frames were temporally displaced by ±1 frame (±42ms at 24fps).

Verification

• Tested on commercial 3D Blu-ray MVC streams (130,000+ frame pairs, High Profile, CABAC, max_num_reorder_frames=4)
• Before fix: periodic right-eye motion spikes (1.4x mean, up to 2.2x) and left-eye POC ordering corruption
• After fix: both views smooth, correct stereo output across entire stream
• 85/85 JVT conformance streams continue to pass

Note

The same_views bitmask is already used correctly in the max_bump calculation at line 1222 and in bump_frame() itself. These two lines were the only places missing it.