Reduce memory and intermediate work in packing and minimization

[kierandidi]

Scoring, repacking, and minimization retain large temporary arrays that can be avoided without changing public APIs or numerical behavior. This combines the performance work from #587, #588 and #589:

• Share the CUDA annealer's sequential quench workspaces; fix empty-pose boundary reads exposed by the expanded regression coverage.
• Construct rotamer topology from real atoms and transfer six forest fields together, avoiding padded host intermediates and repeated transfers.
• Keep circular L-BFGS histories in place and reorder only small matrices/vectors. Make packing benchmark samples reset inputs/RNG outside timing and finish CUDA work before returning.

Matched measurements: CUDA workspace peak falls by 10.10 MiB in the capacity probe; CPU rotamer construction improves from 120.25 to 111.35 ms (one pose) and 337.01 to 310.95 ms (three poses); ragged forest host peak falls from 22.87 to 10.99 MB. L-BFGS helpers improve 1.09–1.23× at histories 32/128, with operator allocation totals falling from 36.9 to 0.80 MB in the large-history probe. History 8 can be about 3% slower (~2 μs). These are component results: full packing/FastRelax timing did not establish an end-to-end speedup. Separate-process packing results were noisy/slower; alternating same-process comparisons were neutral.

Validation: 405 CPU packing tests and 120 CPU optimization tests passed, plus independent recurrence/autograd and real-minimization parity checks. On L40S / CUDA 13 / PyTorch 2.10.0a0, 136 integrated CUDA tests passed; the final empty-pose fix then passed all 31 streaming-graph CUDA tests and 15 focused CPU tests. Seeded comparisons preserve scores, coordinates, assignments, and subsequent RNG draws. The consolidated components exactly match those tested commits; formatting and diff checks pass. GitHub CI and repository review remain the merge gates.

Companion PRs: #583 for environment/release/docs; #590 for cleanup. This PR absorbs #588 and #589.