Add fused BA path

[heiwang1997]

Summary:

• Add a fused CUDA BA path exposed by pipeline.slam.ba.fused.
• Preserve the existing generic PyTorch BA solver as the default and fall back when the graph layout is unsupported.
• Add slam_ext.ba_extended for optional depth masking and intrinsic optimization.
• Add parity tests comparing fused BA against the generic solver on real-run-shaped graph-buffer data, including motion-only, limited disparity, intrinsic optimization, sparse-track combination, and a zero-weight self-edge corner case.

Notes:

• The fused path keeps the DROID kernel's internal 0.001 flow-weight convention guarded in Python.
• The CUDA path now includes dense-disparity damping alongside sensor-depth regularization to match ViPE's generic solver semantics.
• Verbose fused BA logging now reports energy computed from the CUDA path rather than launching an extra Python-side energy recomputation.

Benchmark:

• Video: /home/huangjh/Downloads/lyra2-temple.mp4 with pipeline=default (762 frames processed).
• Common overrides: pipeline.output.save_viz=false, pipeline.post.depth_align_model=null, pipeline.output.save_artifacts=false, pipeline.output.save_slam_map=false; SLAM keyframe depth remains at the pipeline=default setting.
• Command env: CUDA_HOME=/usr/local/cuda-12.8 VIPE_EXT_JIT=1 MAX_JOBS=8.
• Extension JIT was already built (ninja: no work to do), so compile time is not included.

End-to-end wall time:

config	run 1	run 2	mean
pipeline.slam.ba.fused=false	115.74s	116.03s	115.88s
pipeline.slam.ba.fused=true	78.60s	78.20s	78.40s

End-to-end delta: 37.48s saved, 32.3% lower wall time, 1.48x speedup.

Isolated BA timing:

run	ba.fused=false BA time	ba.fused=true BA time	saved	BA speedup
1	45.05s	8.01s	37.04s	5.63x
2	45.39s	7.75s	37.64s	5.86x
mean	45.22s	7.88s	37.34s	5.74x

BA-only method: timed GraphBuffer.bundle_adjustment with CUDA synchronization before and after each call, using the same video and common overrides. Each run made 1723 BA calls. Isolated BA time drops by about 82.6%, and the BA-only saving accounts for essentially all of the measured end-to-end saving on this benchmark.

Validation:

• CUDA_HOME=/usr/local/cuda-12.8 VIPE_EXT_JIT=1 MAX_JOBS=8 uv run python -m unittest tests.test_graph_buffer_fused_ba -v
• CUDA_HOME=/usr/local/cuda-12.8 VIPE_EXT_JIT=1 MAX_JOBS=8 uv run python -m unittest tests.test_slam_ext_ba_extended -v
• uv run pytest tests/test_config.py -q
• uv run ruff format --check vipe/config/slam.py vipe/slam/components/buffer.py tests/test_config.py tests/test_graph_buffer_fused_ba.py tests/test_slam_ext_ba_extended.py
• uv run ruff check vipe/config/slam.py vipe/slam/components/buffer.py tests/test_config.py tests/test_graph_buffer_fused_ba.py tests/test_slam_ext_ba_extended.py
• uv run mypy vipe/config/slam.py vipe/slam/components/buffer.py tests/test_config.py tests/test_graph_buffer_fused_ba.py