[YoussefMadkour]

Fixes #506

Problem

Building the SAM 3.1 predictor with build_sam3_multiplex_video_predictor prints hundreds of Missing keys / Unexpected keys warnings, which makes it look like the checkpoint failed to load. There are two causes:

1. The merged checkpoint is loaded twice. The predictor passes checkpoint_path to build_sam3_multiplex_video_model, which loads the merged checkpoint into the bare tracker. The merged checkpoint's keys are prefixed tracker.model.* / detector.*, while the tracker's own keys are not, so this first load matches nothing. It only prints warnings and reads the multi-GB file an extra time. The same file is then loaded again, correctly, into the full demo model a few lines later.
2. 64 RoPE buffers are reported missing. With use_rope_real=True (the predictor's default), each ViT attention block registers freqs_cis_real / freqs_cis_imag. These are derived from freqs_cis at init and aren't in the checkpoint, which accounts for the Missing keys (64) in the issue.

Fix

• Build the tracker with checkpoint_path=None inside build_sam3_multiplex_video_predictor. Its weights still come from the single load into the demo model.
• Register freqs_cis_real / freqs_cis_imag with persistent=False. They're still created and moved with the module as before, but they're no longer expected in, or saved to, state dicts. freqs_cis itself is unchanged.

The weights the model ends up with don't change: I built the tracker on its own and confirmed none of its 931 state-dict keys start with tracker. or detector., so the first load never loaded anything.

Tests

Added test/test_multiplex_checkpoint_loading.py (no checkpoint or GPU needed):

• build_sam3_multiplex_video_predictor no longer passes the checkpoint to the tracker build (it stops the build at that call with a mock).
• A small ViT with use_rope_real=True keeps the real/imag buffers on the module, leaves them out of state_dict(), and strictly loads a state dict saved without them.

Both tests fail on main and pass with this change.