refactor: reuse sam3_read_f32 for small embedding readbacks

[PABannier]

Six places in sam3.cpp open-code the same readback:

if (t->type == GGML_TYPE_F16) { read into tmp; ggml_fp16_to_fp32_row(...); } else { read F32; }

The tensors are pe_gaussian, the 4 point_embeds, not_a_point_embed, no_mask_embed, no_obj_embed_spatial and no_obj_ptr. sam3_read_f32(t, dst, n) already does exactly this, and it also dequantizes quantized tensors, which the inline copies would have misread as F32. Each site becomes one call.

Stacked on #25 because the no_obj_ptr read sits inside sam3_track_frame, which #25 restructures. Merge #25 first; GitHub will retarget this PR to main. The identical readbacks in the test-API region are left alone, since #20 deletes them.

Verification: a parity tool runs encode, PVS point + box, and a 3-frame visual track, then compares the output byte for byte with origin/main. The prompt-encoder caches and the SAM2.1 no_obj_embed_spatial path are covered.

PARITY OK   edgetam_f16 gpu0
PARITY OK   edgetam_f16 gpu1
PARITY OK   sam2.1_hiera_tiny_q8_0 gpu0
PARITY OK   sam2.1_hiera_tiny_q8_0 gpu1

The no_obj_ptr read in sam3_track_frame (text-prompted SAM3) only has a compile check, because no full SAM3 checkpoint is available locally.

git diff --shortstat: 1 file changed, 7 insertions(+), 49 deletions(-).

Part of an over-engineering cleanup series.

🤖 Generated with Claude Code

https://claude.ai/code/session_01M2qLmQ9ag49hT61V6i9qAq