Improve autotune batch size and CPU count scanning

[scal444]

Autotune now steps in 64 element increments by default, cutting down on the search space. CPU space is now physical core limited by default.

[greptile-apps]

Greptile Summary

This PR tightens the autotune search spaces in two ways: batch sizes are switched from log-uniform ranges to stepped integer ranges (multiples of 64/128), and CPU-thread counts are now bounded by physical core count (read from /proc/cpuinfo) rather than logical count. A new (low, high, step) spec type is plumbed through suggest_from_space and collect_int_from_space in _core.py, with a positive-step guard in the suggestion path.

• Stepped batch size: batchSize for FF, embed, and substructure tunables moves from log-uniform to (low, high, step) tuples (e.g. (64, 1024, 64)), reducing the effective search space while preserving numeric ordering for TPE.
• Physical core detection: _physical_cpu_count_from_proc reads distinct (physical id, core id) pairs from /proc/cpuinfo and falls back to os.cpu_count() when the file is absent or the expected fields are missing (ARM64, some older kernels).
• batchesPerGpu / workerThreads cap: All three tuners now cap the per-GPU thread upper bound at min(8, cpus // num_gpus), reflecting the empirical point of diminishing returns for batched dispatch.

Confidence Score: 5/5

Safe to merge; the logic changes are well-scoped and the new physical-core fallback degrades gracefully.

The stepped-range plumbing is correct and tested end-to-end. The physical CPU detection reads a well-known file, returns None on any failure, and the caller clamps to a floor of 1, so there is no crash path. The batchesPerGpu cap at 8 and the narrowed batchSize ranges are intentional tuning decisions with no correctness risk.

_ff_common.py is worth a second look on ARM64/Grace Hopper targets, where /proc/cpuinfo lacks physical id fields and the function silently falls back to logical count.

Important Files Changed

Filename	Overview
nvmolkit/autotune/_core.py	Adds stepped integer range (low, high, step) support to suggest_from_space (with step <= 0 guard) and collect_int_from_space (snapping to nearest multiple of step from low).
nvmolkit/autotune/_ff_common.py	Introduces _physical_cpu_count_from_proc to read distinct (physical id, core id) pairs from /proc/cpuinfo, falling back to os.cpu_count() if the file is missing or fields are absent; default_ff_search_space switches batchSize to stepped multiples of 64 and caps batchesPerGpu at 8.
nvmolkit/autotune/tune_embed_molecules.py	Mirrors the FF changes: batchSize switched to (64, 1024, 64) stepped range and batchesPerGpu capped at min(8, cpus // num_gpus).
nvmolkit/autotune/tune_substructure.py	Same pattern: batchSize switched to (128, 1024, 128) and workerThreads per-GPU cap now also bounded at 8.
nvmolkit/tests/test_autotune.py	Updated existing tests for the new per-GPU-8 cap and adds new tests for stepped batchSize, the batchesPerGpu 8-cap, and _physical_cpu_count_from_proc SMT deduplication.

_{Reviews (3): Last reviewed commit: "formatting" | Re-trigger Greptile}

[evasnow1992]

LGTM. Only one minor comment. Thanks!