refactor: use sam3_graph_compute in sam2_encode_image_hiera

[PABannier]

sam2_encode_image_hiera open-coded exactly what sam3_graph_compute does: set the CPU thread count when the backend is CPU, run ggml_backend_graph_compute, and check the status. Every other stage in sam3.cpp already goes through the helper. The helper also logs graph compute failed: <status>, so the caller's own error print at this call site is dropped.

Rebased onto main. The earlier conflict was with #27 (RAII contexts/allocators), which had already removed the manual ggml_gallocr_free/ggml_free lines this diff sat on; the failure path is now just return false;.

Verification: a parity tool runs encode, PVS point + box, and a 3-frame visual track, then compares the output byte for byte with main. The SAM2.1 runs go through this function.

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

A Release build with -DSAM3_BUILD_TESTS=ON has no errors or new warnings.

git diff --shortstat main: 1 file changed, 1 insertion(+), 5 deletions(-).

Not fixed here, to keep the diff as reviewed: sam3_encode_vit_from_preprocessed_selective (~sam3.cpp:7128) still open-codes the same pattern, inside the test/debug region that #27 skipped.

Part of an over-engineering cleanup series.

🤖 Generated with Claude Code

https://claude.ai/code/session_01M2qLmQ9ag49hT61V6i9qAq

[PABannier]

Fix the merge conflicts

[PABannier]

Rebased onto main; conflict resolved and force-pushed. GitHub now reports the PR as mergeable.

The conflict was with #27 (RAII contexts/allocators), which had already deleted the manual ggml_gallocr_free(galloc); ggml_free(ctx0); lines this diff sat on top of. Resolution keeps #27's version, so the failure path is just return false;:

    // Compute
    if (!sam3_graph_compute(model.backend, graph, state.n_threads)) {
        return false;
    }

The diff is now 1 insertion, 5 deletions in sam2_encode_image_hiera — the thread-count setup, the status check and the duplicate error print, all of which sam3_graph_compute already does (it also logs the status string).

Verified: encode, PVS point + box, and a 3-frame visual track match main byte for byte on EdgeTAM f16 and SAM2.1-tiny q8_0, CPU and Metal (PARITY OK ×4); the SAM2.1 runs go through this function. A Release build with -DSAM3_BUILD_TESTS=ON has no errors or new warnings.

One note for later, left out to keep this diff as reviewed: sam3_encode_vit_from_preprocessed_selective (~sam3.cpp:7128) still open-codes the same pattern. It's in the test/debug region that #27 skipped, so it kept its manual frees and needs a slightly different edit.